{ "@context": "https://schema.org", "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-05-20-qwen3-7-the-agent-frontier/#webpage", "@type": [ "WebPage", "ArchiveComponent" ], "articleBody": "![Qwen3.7 Main Image](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/qwen3.7-max-banner.png)\n\nToday we introduce **Qwen3.7-Max**, our latest proprietary model designed for the agent era. Qwen3.7-Max is built to be a versatile agent foundation — equally capable of writing and debugging code, automating office workflows, and sustaining autonomous execution across hundreds or thousands of steps.\n\nWhat sets Qwen3.7-Max apart is the breadth and depth of its agent capabilities. It excels as a coding agent, from frontend prototyping to complex multi-file engineering. It serves as a reliable office and productivity assistant through MCP integrations and multi-agent orchestration. It sustains coherent reasoning across extremely long horizons — as demonstrated by a 35-hour, fully autonomous kernel optimization run comprising over 1,000 tool calls. It generalizes across agent scaffolds, performing consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or other frameworks.\n\n- **Qwen3.7-Max** — now available via [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/):\n - frontier coding agent: from frontend prototyping to complex software engineering\n - office productivity and workflow automation via MCP and multi-agent orchestration\n - sustained autonomous execution across long-horizon tasks\n - cross-scaffold generalization across diverse agent frameworks\n- Call via API on [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/).\n\n## Performance\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/Qwen3.7-Max-Score.png)\n\n| | Opus-4.6 Max | K2.6 Thinking | GLM-5.1 Thinking | DS-V4-Pro Max | Qwen3.6-Plus | Qwen3.7-Max |\n| --------------------------- | ------------ | ------------- | ---------------- | ------------- | ------------ | ----------- |\n| Coding Agent | | | | | | |\n| Terminal Bench 2.0-Terminus | 65.4 | 66.7 | 63.5 | 67.9 | 61.6 | 69.7 |\n| SWE-Verified | 80.8 | 80.2 | -- | 80.6 | 78.8 | 80.4 |\n| SWE-Pro | 57.3 | 59.5 | 58.8 | 59.0 | 56.6 | 60.6 |\n| SWE-Multilingual | 77.5 | 76.7 | -- | 76.2 | 73.8 | 78.3 |\n| NL2repo | 47.6 | 42.8 | 41.0 | 35.5 | 34.4 | 47.2 |\n| SciCode | 51.9 | 52.2 | 45.1 | -- | 41.4 | 53.5 |\n| QwenWebDev | 1617 | -- | 1564 | 1570 | 1500 | 1568 |\n| QwenSVG | 1541 | 1325 | 1605 | 1506 | 1432 | 1608 |\n| General Agent | | | | | | |\n| Qwenclaw | 65.5 | 54.7 | 58.7 | 59.2 | 57.2 | 64.3 |\n| CoWorkBench | 68.2 | 58.2 | 66.0 | 66.3 | 64.5 | 67.2 |\n| ClawEval | 70.4 | 61.5 | 62.7 | 58.4 | 57.1 | 65.2 |\n| Skillsbench | -- | 56.2 | 53.1 | 52.3 | 45.7 | 59.2 |\n| BFCL-V4 | 76.7 | 71.3 | 70.9 | 70.6 | 68.9 | 75.0 |\n| MCP-Mark | 56.7 | 55.9 | 57.5 | 57.1 | 48.2 | 60.8 |\n| MCP-Atlas | 75.8 | 66.6 | 71.8 | 73.6 | 74.1 | 76.4 |\n| Vitabench | -- | 39.1 | 45.1 | 51.9 | 42.8 | 47.9 |\n| SpreadSheetBench-v1 | 89.3 | 84.5 | 85.2 | 84.9 | 80.2 | 87.0 |\n| Kernel Bench L3 | 2.63/98% | 1.41/80% | 2.00/78% | 1.07/54% | 1.03/48% | 1.98/96% |\n| HLE w/ tools | 53.0 | 54.0 | 52.3 | 48.2 | 50.2 | 53.5 |\n| QwenWorldBench | 56.1 | 50.9 | 50.2 | 52.3 | 47.6 | 57.3 |\n| STEM & Reasoning | | | | | | |\n| GPQA Diamond | 91.3 | 90.5 | 86.2 | 90.1 | 90.4 | 92.4 |\n| HLE | 40.0 | 36.4 | 34.7 | 37.7 | 28.8 | 41.4 |\n| LiveCodeBench | 88.8 | 89.6 | -- | 93.5 | 87.1 | 91.6 |\n| HMMT 2026 Feb | 96.2 | 92.7 | 89.4 | 95.2 | 87.8 | 97.1 |\n| IMOAnswerBench | 75.3 | 86.0 | 83.8 | 89.8 | 83.8 | 90.0 |\n| CritPT | 12.6 | 8.0 | 4.6 | 12.9 | 2.9 | 11.4 |\n| Apex | 34.5 | 24.0 | 11.5 | 38.3 | 8.8 | 44.5 |\n| General Capability | | | | | | |\n| MMLU-Pro | 89.7 | 87.1 | 86.3 | 87.5 | 88.5 | 89.6 |\n| MMLU-Redux | 95.2 | 95.3 | 94.3 | 94.8 | 94.5 | 95.0 |\n| SuperGPQA | 72.5 | 71.3 | 68.0 | 69.9 | 71.6 | 73.6 |\n| IFEval | 91.9 | 94.5 | 94.5 | 91.9 | 94.3 | 94.3 |\n| IFBench | 62.5 | 76.0 | 76.0 | 77.0 | 74.2 | 79.1 |\n| MRCR-v2 128k | 84.0 | 63.1 | 62.0 | 74.4 | 85.9 | 90.4 |\n| Multilingualism | | | | | | |\n| WMT24++ | 82.7 | 81.6 | 81.8 | 82.2 | 84.3 | 85.8 |\n| MAXIFE | 81.3 | 87.7 | 87.7 | 88.9 | 88.2 | 89.2 |\n| MMMLU | 90.6 | 87.5 | 87.2 | 87.9 | 89.5 | 90.3 |\n| MMLU-ProX | 86.1 | 83.7 | 83.9 | 83.9 | 84.7 | 87.0 |\n| NOVA-63 | 59.1 | 56.7 | 54.6 | 52.8 | 57.9 | 59.0 |\n| INCLUDE | 87.4 | 84.2 | 84.3 | 86.1 | 85.1 | 86.2 |\n| Global PIQA | 91.2 | 89.2 | 89.5 | 90.5 | 89.8 | 91.4 |\n| PolyMATH | 80.2 | 82.7 | 67.6 | 72.0 | 77.4 | 86.5 |\n\n\\* Terminal-Bench 2.0: Harbor/Terminus-2 harness; 5h timeout, 12 CPU/24 GB RAM; temp=1.0, top\\_p=0.95, top\\_k=20, max\\_tokens=80K, 256K ctx; avg of 5 runs. All experiments prepend a token at each turn, allowing the model to decide whether to engage extended thinking.\\\n\\* SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top\\_p=0.95, 200K context window.\\\n\\* SWE-bench Pro: Problematic tasks corrected and all baselines evaluated on the refined benchmark.\\\n\\* NL2Repo: Evaluated via Claude-code. We disable Bash commands that attempt to access the specific repository, such as pip download, pip install, and git clone.\\\n\\* QwenWebDev: Internal front-end code generation benchmark; bilingual (EN/CN), 7 categories; auto-render + multimodal judge; BT/Elo rating.\\\n\\* QwenClawBench: a real-user-distribution Claw agent benchmark; open-source: [https://github.com/SKYLENAGE-AI/QwenClawBench](https://github.com/SKYLENAGE-AI/QwenClawBench).\\\n\\* CoWorkBench: an internal cowork benchmark; long-horizon tasks across computer science, finance, law, medical, and other productivity domains.\\\n\\* SkillsBench: Evaluated via OpenCode on 78 tasks (excluding 9 external API-dependent tasks); avg of 5 runs.\\\n\\* MCP-Mark: GitHub MCP v0.30.3; Playwright responses truncated at 32K tokens.\\\n\\* MCP-Atlas: Public set score; gemini-2.5-pro judger.\\\n\\* VITA-Bench: Avg subdomain scores; using claude-4.5-sonnet as judger, as the older official judgers are no longer available.\\\n\\* Kernel Bench L3: Metrics reported: median of per-problem speedup over PyTorch eager reference / fraction of problems faster than torch.compile, across 50 problems. Each test sample runs in an isolated Docker container with one H100 80GB GPU, with internet access restricted to the CUTLASS codebase and official CUDA documentation, limited to 500 tool calls with early stopping after 100 non-improving turns. GPT-5.4 (xhigh) is applied to detect potential hacking behaviors. CUPTI is used for kernel-level timing.\\\n\\* QwenWorldBench: Internal benchmark for evaluating LLMs as world models for simulating agentic environments; 7 domains (Terminal, SWE, MCP, Search, OS, Android, Web); open-ended 5-dim rubric judge grounded in real-environment feedback.\\\n\\* Reasoning scenarios: Recommended system prompt: \"Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.\"\\\n\\* MRCR-v2: 128K context subset containing 8 needles utilized; evaluation protocol adopted from https://github.com/google-deepmind/eval_hub/tree/master/eval_hub/mrcr_v2.\\\n\\* WMT24++: Harder WMT24 subset; avg scores on 55 langs via XCOMET-XXL.\\\n\\* MAXIFE: Accuracy on EN + multilingual prompts (23 settings total).\\\n\\* MMLU-ProX: Avg accuracy across 29 languages.\\\n\\* Empty cells (--) indicate scores not yet available.\n\nIn **coding agents**, Qwen3.7-Max performs strongly on SWE-Pro (60.6), SWE-Multilingual (78.3), SciCode (53.5), and QwenSVG (1608). On Terminal Bench 2.0-Terminus (69.7), it outperforms DS-V4-Pro Max (67.9). On SWE-Verified (80.4), it is on par with Opus-4.6 Max (80.8) and DS-V4-Pro Max (80.6).\n\nIn **general-purpose agents**, improvements are even more pronounced. Qwen3.7-Max performs exceptionally well on MCP-Mark (60.8 vs. GLM-5.1’s 57.5), MCP-Atlas (76.4 vs. Opus-4.6’s 75.8), and Skillsbench (59.2 vs. K2.6’s 56.2), and demonstrates strong GPU kernel optimization capabilities on Kernel Bench L3 (1.98x median speedup, 96% win rate). It also scores highly on BFCL-V4 (75.0), Qwenclaw (64.3), and ClawEval (65.2), closely approaching Opus-4.6 Max. On the office automation benchmark SpreadSheetBench-v1, it achieves a top-tier score of 87.\n\nIn **reasoning**, Qwen3.7-Max achieves leading results on GPQA Diamond (92.4 vs. Opus-4.6’s 91.3), HLE (41.4 vs. Opus-4.6’s 40), HMMT 2026 Feb (97.1 vs. Opus-4.6’s 96.2), IMOAnswerBench (90 vs. DS-V4-Pro’s 89.8), and Apex (44.5 vs. DS-V4-Pro’s 38.3), demonstrating exceptional strength on the hardest reasoning benchmarks.\n\nIn **general capabilities and multilingualism**, Qwen3.7-Max stands out on IFBench (79.1 vs. DS-V4-Pro’s 77.0), demonstrating precise instruction following. It achieves leading scores on WMT24++ (85.8) and MAXIFE (89.2), confirming top-tier multilingual understanding and translation quality. It also delivers strong results on SuperGPQA (73.6) and QwenWorldBench (57.3).\n\nNotably, these scores are drawn from a wide variety of agent scaffolds. Rather than optimizing for any single framework, Qwen3.7-Max delivers consistently across Claude Code, OpenClaw, Qwen Code, and custom tool-use frameworks, making it a reliable drop-in backbone for any agent system.\n\n## Cowork Productivity Assistant\n\nQwen3.7-Max serves as your advanced coworker for real-world productivity. Its powerful agent capabilities fundamentally streamline professional workflows — synthesizing complex information, performing in-depth data analysis and modeling, and generating publication-ready documents and visualizations — to reliably handle high-complexity enterprise workloads.\n\nQwen3.7-Max features native compatibility with mainstream agent harnesses. For long-horizon tasks, it supports autonomous planning and continuous execution across multi-hour sessions. Through thousands of tool calls and dozens of refinement iterations, it steadily improves output quality. Complex projects that typically require one to two weeks of specialized team effort can now be completed end-to-end within hours, delivering measurable productivity gains.\n\n## Agent Scaling\n\nBuilding on the environment scaling approach introduced in Qwen3.5, we have continued to aggressively expand both the quality and diversity of agentic training environments in Qwen3.7. Just as language models generalize from diverse pretraining text, we find that agentic capabilities generalize from diverse training environments.\n\nAs shown in the figure below, this environment scaling produces a clear and consistent improvement trajectory, with Qwen3.7-Max achieving a top-3 average ranking that approaches Claude-4.6-Opus-Max. Crucially, all benchmarks in our evaluation feature entirely unseen, out-of-domain environments that were never present in training.\n\nWe also observe a striking predictability in the scaling behavior: performance gains across any subset of benchmarks are highly consistent and can reliably predict the relative gains on the remaining benchmarks or the overall average, suggesting that environment scaling drives genuine capability generalization rather than benchmark-specific improvement. Further analysis of the scaling dynamics and methodology will be detailed in our upcoming technical report.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/agent_scaling.png#center)\n\n## Cross-Harness Generalization\n\nOur Rollout environment infrastructure decouples each training instance into three orthogonal components — Task, Harness, and Verifier — that can be freely recombined. We support a wide range of harnesses and their evolving versions, and ground our environments in real-world settings rather than synthetic proxies. This decoupled design enables combinatorial scaling: the same task is paired with diverse harnesses (across types and versions) and verifiers at minimal marginal cost. More critically, it enables cross-harness and cross-verifier RL training, where the model encounters identical tasks under varying harness configurations, forcing it to learn generalizable problem-solving strategies rather than harness-specific shortcuts. Across QwenClawBench and CoWorkBench, Qwen3.7-Max delivers strong, consistent performance regardless of the harness used at evaluation time, confirming that the model has learned to solve tasks — not to exploit particular harnesses.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/harness-generalization.png#center)\n\n## Self-Evolving in the Wild\n\n[Extend Attention](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/layers/attention/triton_ops/extend_attention.py) is a production-grade, variable-length multi-head attention operator in SGLang. In our test scenario, it computes attention scores between newly generated tokens and a prefix KV-cache of up to 32K entries with MTP — a memory-bound, latency-critical kernel in LLM serving. The reference implementation is SGLang’s official Triton implementation.\n\nWe tasked Qwen3.7-Max with optimizing this kernel on an ECS instance equipped with T-Head ZW-M890 PPUs — a hardware platform **never seen during training**. The model had no prior profiling data, no hardware documentation, and no example kernels for this architecture. It started from an empty workspace containing only a task description, the existing SGLang implementation, and an evaluation script.\n\nOver the course of **\\~35 hours of continuous autonomous execution**, the model performed **432 kernel evaluations across 1,158 tool calls**. It wrote, compiled, profiled, and iteratively improved the Extend Attention Kernel entirely on its own — diagnosing compilation failures, fixing correctness bugs, identifying performance bottlenecks through runtime profiling, and redesigning the kernel architecture multiple times.\n\nThe final result: **10.0x geometric mean speedup** over the Triton reference, measured across multiple workloads. The optimization trajectory shows sustained, non-trivial progress far beyond the first few hours: the model was still finding meaningful improvements after 30+ hours, demonstrating that long-horizon autonomous optimization is not just feasible but productive.\n\n[docs.qwenlm.ai](https://docs.qwenlm.ai/resources/hUbVe_evolving_ppu_case.html)\n\nKey structural transitions in the optimization trajectory\n\n1. **Split-KV parallelism** (0.33x → 2.58x, ~2h): The initial kernel launched only 8 blocks (4 tokens × 2 KV heads × 1 batch) on 36 SMs, leaving most SMs idle. The model redesigned the kernel with Split-KV partitioning — dividing the prefix KV-cache across multiple thread blocks per query — and introduced a separate reduction kernel using online softmax rescaling to merge partial results.\n\n2. **Launch and allocation overhead removal** (2.58x → 5.37x, ~2.5h): The model systematically removed host-device synchronization overhead: replacing per-call `cudaMalloc`/`cudaFree` with pre-allocated `torch::empty` tensors, eliminating synchronous `cudaMemcpy` calls for prefix length queries by using tensor metadata instead, and unrolling the inner loop 2x to amortize loop control overhead and increase instruction-level parallelism.\n\n3. **Workload-adaptive split tuning** (5.37x → 6.85x, ~3h): The model evolved from a fixed split divisor to workload-size-dependent heuristics — applying more aggressive splitting for smaller inputs and tuning per-workload split counts to maximize SM wave occupancy on the 36-SM architecture.\n\n4. **Reduction and batching optimizations** (6.85x → 8.50x, 3h–25h): Eliminating shared memory barriers by switching to register-based K/V loading for higher SM occupancy, persistent static tensors for partial results to avoid per-call allocation, more aggressive split heuristics for small inputs, and a batched softmax update (4 `expf` calls instead of 6) to reduce per-token overhead. Q pre-scaling by `sm_scale` eliminated a per-iteration floating-point multiply after warp reduction.\n\n5. **MTP γ=4 specialized kernel** (8.50x → 10.0x, 32h–35h): The most significant architectural redesign — restructuring the kernel to process all 4 query tokens simultaneously per block, sharing K/V loads across queries to amortize memory access cost. Combined with `__ldg` read-only cache intrinsics for V buffer loads, multi-query batched attention output reduction, register pressure tuning, and re-tuned split heuristics, this delivered the final ~1.2x improvement in the last hours.\n\nWe also ran the same task with several other models under identical conditions. GLM 5.1 reached 7.3x; Kimi K2.6 reached 5.0x; DeepSeek V4 Pro reached 3.3x; Qwen3.6-Plus reached 1.1x. Models that stopped early did so because the agent issued no tool calls for five consecutive rounds — the model concluded it could no longer make progress and voluntarily ended the session.\n\nIn addition to achieving strong kernel generation results on PPUs, Qwen3.7-Max also generates high-quality, production-grade kernels across a variety of NVIDIA GPUs. For example, on KernelBench L3, Qwen3.7-Max is able to produce accelerated kernels for 96% of the scenarios, compared to 98% for Opus-4.6, 78% for GLM 5.1, 80% for Kimi K2.6, 54% for DeepSeek V4 Pro, and 48% for Qwen3.6-Plus.\n\nThis result highlights two properties of Qwen3.7-Max as a foundation model powering long-horizon autonomous agents: **sustained long-horizon reasoning** — the model maintains coherent optimization strategy across over a thousand tool calls without losing context or regressing — and **strong in-context generalization** — it produces competitive kernels for an architecture it has never encountered, relying on runtime feedback rather than memorized hardware knowledge.\n\n## Reward Hacking Monitoring for Long-Horizon Training\n\nWe integrated Qwen3.7-Max into the Reinforcement Learning (RL) monitoring for Software Engineering (SWE) tasks, successfully building a framework for reward hacking self-monitoring and rule self-evolution. During RL experiments exceeding 80 hours, the model autonomously retrieved and replayed training trajectories, executing over 10,000 calls. The system systematically identified candidate hacking patterns (such as attempts to bypass constraints to access ground-truth answers on GitHub) while performing rule verification, counter-example mining, and iterative optimization.\n\nAs a result, Qwen3.7-Max achieved multiple rounds of rule self-evolution, adding 13 new heuristic rules and accurately flagging 1,618 hacking cases. This not only ensured the stability of RL rewards but also facilitated the continuous self-improvement of the model as a sophisticated software engineering agent.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/autonomous_hacking_detect.png#center)\n\n## Long-Horizon Planning and Execution in Startup Management\n\nWithin the framework of Dynamic Cumulative Survival Games, we have scaled the temporal complexity of training tasks to specifically reinforce long-horizon planning and execution capabilities. This advancement enhances the agent’s policy consistency throughout sequential decision-making trajectories exceeding a thousand steps, enabling it to continuously construct hypotheses, dynamically adjust strategies based on environmental feedback, and accumulate long-term experience and memory. Consequently, the agent maintains a stable execution cadence over vast time horizons, remaining resilient to the common pitfalls of context rot and instruction drift.\n\nIn YC-Bench — a benchmark simulating the full year-long lifecycle of a startup — the agent must navigate hundreds of decision-making rounds ranging from personnel management and contract screening to malicious client identification, all while maintaining a profit margin against rising labor costs. Qwen3.7-Max achieved a total revenue of 2.08M USD, which is double the performance of Qwen3.6-Plus (1.05M USD) and 5.9 times that of Qwen3.5-Plus (352K USD), successfully completing 237 tasks. Beyond the metrics, the model demonstrated a profound capacity for strategic evolution across context windows: it actively explored potential clients, identified and blacklisted malicious traps, prioritized reliable revenue streams, and autonomously recovered from mid-term crises to eventually converge into a stable, high-efficiency execution loop.\n\n[docs.qwenlm.ai](https://docs.qwenlm.ai/resources/YdJUk_yc_bench.html)\n\n## Build with Qwen3.7\n\nQwen3.7-Max is now available through [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/). You can integrate it with popular agent frameworks and coding assistants.\n\n### API Usage\n\nQwen3.7-Max supports the `preserve_thinking` feature: preserving thinking content from all preceding turns in messages, which is **recommended for agentic tasks**.\n\n#### Alibaba Cloud Model Studio\n\nAlibaba Cloud Model Studio supports industry-standard protocols, including chat completions and responses APIs compatible with OpenAI’s specification, as well as an API interface compatible with Anthropic.\n\n```python\n\"\"\"\nEnvironment variables:\n DASHSCOPE_API_KEY: Your API Key from https://modelstudio.console.alibabacloud.com\n DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.\n - Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1\n - Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1\n - US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1\n\"\"\"\nfrom openai import OpenAI\nimport os\n\napi_key = os.environ.get(\"DASHSCOPE_API_KEY\")\nif not api_key:\n raise ValueError(\n \"DASHSCOPE_API_KEY is required. \"\n \"Set it via: export DASHSCOPE_API_KEY='your-api-key'\"\n )\n\nclient = OpenAI(\n api_key=api_key,\n base_url=os.environ.get(\n \"DASHSCOPE_BASE_URL\",\n \"https://dashscope-intl.aliyuncs.com/compatible-mode/v1\",\n ),\n)\n\nmessages = [{\"role\": \"user\", \"content\": \"Write a Python function to merge two sorted linked lists.\"}]\n\ncompletion = client.chat.completions.create(\n model=\"qwen3.7-max\",\n messages=messages,\n extra_body={\n \"enable_thinking\": True,\n # \"preserve_thinking\": True,\n },\n stream=True\n)\n\nreasoning_content = \"\"\nanswer_content = \"\"\nis_answering = False\nprint(\"\\n\" + \"=\" * 20 + \"Reasoning\" + \"=\" * 20 + \"\\n\")\n\nfor chunk in completion:\n if not chunk.choices:\n print(\"\\nUsage:\")\n print(chunk.usage)\n continue\n\n delta = chunk.choices[0].delta\n\n if hasattr(delta, \"reasoning_content\") and delta.reasoning_content is not None:\n if not is_answering:\n print(delta.reasoning_content, end=\"\", flush=True)\n reasoning_content += delta.reasoning_content\n\n if hasattr(delta, \"content\") and delta.content:\n if not is_answering:\n print(\"\\n\" + \"=\" * 20 + \"Answer\" + \"=\" * 20 + \"\\n\")\n is_answering = True\n print(delta.content, end=\"\", flush=True)\n answer_content += delta.content\n```\n\nFor more information, please visit the [API doc](https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?type=model&url=2840915).\n\n### Frontend Coding\n\nQwen3.7-Max can generate rich interactive web applications from a single prompt — including Three.js 3D scenes, Canvas animations, full page layouts, and dynamic SVG.\n\nGesture Controlled Particles System Next\n\nUser\n\n用Three.js创建一个实时交互的3D粒子系统网页。要求：1.通过摄像头检测手掌张合控制粒子群的收缩与扩散，当手掌张开时例子扩散，当手掌握紧时例子收缩为一个球；2.当手势为1时，粒子组成文字（hello, world），当手势为2时组成文字 （I’am Qwen）；3.粒子需实时响应手势变化；4.文字应有3D旋转效果；5. 用html实现\n\nQwen3.7-Max\n\nFashion Magazine Page with Video-Generation Next\n\nUser\n\nBuild a “MONOLITH” luxury fashion magazine cover as an interactive web page. You have access to a video generation script at scripts/dashscope-video-gen.sh that can generate AI videos.\n\nSTEP 1 — Generate 3 custom videos by running these commands: bash scripts/dashscope-video-gen.sh “Extreme close-up cinematic portrait of a high fashion model, dramatic side lighting, black and white with crimson red lipstick, wind blowing hair slowly, Vogue cover aesthetic, shallow depth of field” assets/demo-videos/monolith-hero.mp4 9:16 5\n\nbash scripts/dashscope-video-gen.sh “Cinematic slow-motion haute couture fashion runway walk, dramatic chiaroscuro spotlight, silhouette emerging from darkness, editorial magazine film look with grain” assets/demo-videos/monolith-knockout.mp4 9:16 5\n\nbash scripts/dashscope-video-gen.sh “Overhead cinematic shot of luxury fashion magazine open pages being slowly turned, golden jewelry and perfume bottle on black marble, dramatic spotlight, editorial still life” assets/demo-videos/monolith-editorial.mp4 16:9 5\n\nSTEP 2 — Build the web page using the generated videos: HERO: Full-viewport portrait video (monolith-hero.mp4 as muted autoplay looping background) with duotone overlay. Knockout headline “MONOLITH” revealing monolith-knockout.mp4 playing through transparent text. EDITORIAL: Below-fold split layout with monolith-editorial.mp4 in a magazine-page container with page-curl hover effect, and pull-quotes with architectural line accents. DESIGN: Monochrome + crimson accent, grain overlay, minimal masthead, scroll-triggered animations, sticky top bar with issue number and date.\n\nDo not ask follow-up questions. Do not wait for user input. Pick reasonable defaults and complete the artifact. Write all code to a single index.html file. Run the video generation commands first, wait for them to complete, then build the HTML.\n\nQwen3.7-Max\n\nInstancing Dynamic Next\n\nUser\n\nI want to build a Three.js WebGL page that shows instancing dynamic. Can you make it a single standalone HTML file that works when I just open it in a browser?\n\nQwen3.7-Max\n\n3D Racing Game Next\n\nUser\n\n创建一个3D赛车游戏，补充自动玩游戏的auto模式, 优化赛车游戏的轨道，障碍物，另外请补充金币等机制。\n\nQwen3.7-Max\n\nDragon Boat Next\n\nUser\n\n生成一段 SVG 代码，用于表现一幅充满动感的龙舟竞渡矢量插画。画面的主体是一艘细长的龙舟，船上坐满了划手，他们身穿鲜亮的黄绿色上衣和白色帽子，整齐划一地挥动船桨。船桨切入浑浊的棕色河水中，激起白色水花。船尾站着一名指挥者，穿着色彩鲜艳的短裤；船头飘扬着一面蓝色旗帜，上面有黄色数字“15”。背景中，一座巨大的灰色混凝土桥横跨河面，远处在淡蓝色天空下可见朦胧的城市建筑和绿色树木。整体风格采用现代扁平化设计，色彩鲜明、轮廓清晰。请添加流畅的动画，以表现以下动态效果：划手整齐同步的划桨动作、龙舟轻微的上下起伏、旗帜随风飘动，以及水面的涟漪，从而营造出紧张激烈、充满能量的竞赛氛围。\n\nQwen3.7-Max\n\n### Office Assistant\n\nQwen3.7-Max can act as an intelligent office assistant through tool integration. In this example, it reads a university thesis formatting specification and automatically reformats a messy draft — fixing page layout, heading styles, fonts, margins, table of contents, and reference formatting — all through autonomous office-cli tool calls. *(The sample thesis is AI-generated for demonstration purposes.)*\n\nThesis Formatting with Office Tools Next\n\nTo facilitate front-end display, the original Word document is specifically displayed here as a PDF.\n\nUser\n\n请完成一个论文格式修复任务。 ## 输入文件 - 格式规范说明文件: 研究生学位论文格式规范.docx - 格式混乱版论文（待修复）: 论文\\_格式混乱版.docx ## 输出文件 - 论文\\_格式修复版.docx\n\nWorkspace\n\n[研究生学位论文格式规范.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/%E7%A0%94%E7%A9%B6%E7%94%9F%E5%AD%A6%E4%BD%8D%E8%AE%BA%E6%96%87%E6%A0%BC%E5%BC%8F%E8%A7%84%E8%8C%83.docx)\n\n[论文\\_格式混乱版.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/3.%E8%AE%BA%E6%96%87_%E6%A0%BC%E5%BC%8F%E6%B7%B7%E4%B9%B1%E7%89%88.docx)\n\nQwen3.7-Max\n\n[论文\\_格式修复版.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/%E8%AE%BA%E6%96%87_%E6%A0%BC%E5%BC%8F%E4%BF%AE%E5%A4%8D%E7%89%88_cli.docx)\n\n### LLM-Powered Phyicial-World Navigation Agent\n\nOne more thing, Qwen3.7-Max now can operate a robot dog through tool-use calls — performing physical understanding, planning, memory and decision-making in physical environments, powered by our robotics agent harness Qwen-RobotClaw, navigation foundation model Qwen-RobotNav, and several vision tools built with Qwen-plus model. In the demo below, the left panel shows a 20 mins agent’s tool-call interaction flow in the physical-world; the center shows the quadruped robot’s first-person view along its trajectory, and the right shows the agent’s long-term memory.\n\n### Coding Assistants\n\nQwen3.7-Max integrates seamlessly with popular agent frameworks and coding assistants:\n\n#### Claude Code\n\nQwen APIs support the Anthropic API protocol, enabling direct use with **Claude Code**:\n\n```bash\nnpm install -g @anthropic-ai/claude-code\n\nexport ANTHROPIC_MODEL=\"qwen3.7-max\"\nexport ANTHROPIC_SMALL_FAST_MODEL=\"qwen3.7-max\"\nexport ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic\nexport ANTHROPIC_AUTH_TOKEN=\n\nclaude\n```\n\n#### OpenClaw\n\nConnect to [OpenClaw](https://openclaw.ai) via [Model Studio](https://www.alibabacloud.com/help/en/model-studio/openclaw):\n\n```bash\ncurl -fsSL https://molt.bot/install.sh | bash\nexport DASHSCOPE_API_KEY=\nopenclaw dashboard\n```\n\nConfigure `~/.openclaw/openclaw.json`:\n\n```json\n{\n \"models\": {\n \"mode\": \"merge\",\n \"providers\": {\n \"modelstudio\": {\n \"baseUrl\": \"https://dashscope-intl.aliyuncs.com/compatible-mode/v1\",\n \"apiKey\": \"DASHSCOPE_API_KEY\",\n \"api\": \"openai-completions\",\n \"models\": [\n {\n \"id\": \"qwen3.7-max\",\n \"name\": \"qwen3.7-max\",\n \"reasoning\": true,\n \"input\": [\"text\"],\n \"contextWindow\": 1000000,\n \"maxTokens\": 65536\n }\n ]\n }\n }\n },\n \"agents\": {\n \"defaults\": {\n \"model\": {\n \"primary\": \"modelstudio/qwen3.7-max\"\n }\n }\n }\n}\n```\n\n#### Qwen Code\n\n[Qwen Code](https://qwen.ai/qwencode) is deeply optimized for the Qwen series:\n\n```bash\nnpm install -g @qwen-code/qwen-code@latest\nqwen\n```\n\n## Summary\n\nQwen3.7-Max is our most versatile and capable model for agent-driven workflows. From coding and office automation to long-horizon autonomous tasks, it combines frontier-level reasoning with robust cross-scaffold generalization and the ability to sustain productive execution over extended periods — providing a powerful foundation for building the next generation of AI agents. We welcome community feedback and look forward to seeing what you build.\n\n## Citation\n\n```bibtex\n@misc{qwen37,\n title = {{Qwen3.7}: The Agent Frontier},\n url = {https://qwen.ai/blog?id=qwen3.7},\n author = {{Qwen Team}},\n month = {May},\n year = {2026}\n}\n```\n", "authors": [ "QwenTeam" ], "captured_at": "2026-10-19T01:26:53.998049939+00:00", "category": "ai", "content_html": "

\"Qwen3.7

\n

Today we introduce Qwen3.7-Max, our latest proprietary model designed for the agent era. Qwen3.7-Max is built to be a versatile agent foundation — equally capable of writing and debugging code, automating office workflows, and sustaining autonomous execution across hundreds or thousands of steps.

\n

What sets Qwen3.7-Max apart is the breadth and depth of its agent capabilities. It excels as a coding agent, from frontend prototyping to complex multi-file engineering. It serves as a reliable office and productivity assistant through MCP integrations and multi-agent orchestration. It sustains coherent reasoning across extremely long horizons — as demonstrated by a 35-hour, fully autonomous kernel optimization run comprising over 1,000 tool calls. It generalizes across agent scaffolds, performing consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or other frameworks.

\n\n

Performance

\n

\"\"

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Opus-4.6 MaxK2.6 ThinkingGLM-5.1 ThinkingDS-V4-Pro MaxQwen3.6-PlusQwen3.7-Max
Coding Agent
Terminal Bench 2.0-Terminus65.466.763.567.961.669.7
SWE-Verified80.880.2--80.678.880.4
SWE-Pro57.359.558.859.056.660.6
SWE-Multilingual77.576.7--76.273.878.3
NL2repo47.642.841.035.534.447.2
SciCode51.952.245.1--41.453.5
QwenWebDev1617--1564157015001568
QwenSVG154113251605150614321608
General Agent
Qwenclaw65.554.758.759.257.264.3
CoWorkBench68.258.266.066.364.567.2
ClawEval70.461.562.758.457.165.2
Skillsbench--56.253.152.345.759.2
BFCL-V476.771.370.970.668.975.0
MCP-Mark56.755.957.557.148.260.8
MCP-Atlas75.866.671.873.674.176.4
Vitabench--39.145.151.942.847.9
SpreadSheetBench-v189.384.585.284.980.287.0
Kernel Bench L32.63/98%1.41/80%2.00/78%1.07/54%1.03/48%1.98/96%
HLE w/ tools53.054.052.348.250.253.5
QwenWorldBench56.150.950.252.347.657.3
STEM & Reasoning
GPQA Diamond91.390.586.290.190.492.4
HLE40.036.434.737.728.841.4
LiveCodeBench88.889.6--93.587.191.6
HMMT 2026 Feb96.292.789.495.287.897.1
IMOAnswerBench75.386.083.889.883.890.0
CritPT12.68.04.612.92.911.4
Apex34.524.011.538.38.844.5
General Capability
MMLU-Pro89.787.186.387.588.589.6
MMLU-Redux95.295.394.394.894.595.0
SuperGPQA72.571.368.069.971.673.6
IFEval91.994.594.591.994.394.3
IFBench62.576.076.077.074.279.1
MRCR-v2 128k84.063.162.074.485.990.4
Multilingualism
WMT24++82.781.681.882.284.385.8
MAXIFE81.387.787.788.988.289.2
MMMLU90.687.587.287.989.590.3
MMLU-ProX86.183.783.983.984.787.0
NOVA-6359.156.754.652.857.959.0
INCLUDE87.484.284.386.185.186.2
Global PIQA91.289.289.590.589.891.4
PolyMATH80.282.767.672.077.486.5
\n

* Terminal-Bench 2.0: Harbor/Terminus-2 harness; 5h timeout, 12 CPU/24 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs. All experiments prepend a token at each turn, allowing the model to decide whether to engage extended thinking.
\n* SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window.
\n* SWE-bench Pro: Problematic tasks corrected and all baselines evaluated on the refined benchmark.
\n* NL2Repo: Evaluated via Claude-code. We disable Bash commands that attempt to access the specific repository, such as pip download, pip install, and git clone.
\n* QwenWebDev: Internal front-end code generation benchmark; bilingual (EN/CN), 7 categories; auto-render + multimodal judge; BT/Elo rating.
\n* QwenClawBench: a real-user-distribution Claw agent benchmark; open-source: https://github.com/SKYLENAGE-AI/QwenClawBench.
\n* CoWorkBench: an internal cowork benchmark; long-horizon tasks across computer science, finance, law, medical, and other productivity domains.
\n* SkillsBench: Evaluated via OpenCode on 78 tasks (excluding 9 external API-dependent tasks); avg of 5 runs.
\n* MCP-Mark: GitHub MCP v0.30.3; Playwright responses truncated at 32K tokens.
\n* MCP-Atlas: Public set score; gemini-2.5-pro judger.
\n* VITA-Bench: Avg subdomain scores; using claude-4.5-sonnet as judger, as the older official judgers are no longer available.
\n* Kernel Bench L3: Metrics reported: median of per-problem speedup over PyTorch eager reference / fraction of problems faster than torch.compile, across 50 problems. Each test sample runs in an isolated Docker container with one H100 80GB GPU, with internet access restricted to the CUTLASS codebase and official CUDA documentation, limited to 500 tool calls with early stopping after 100 non-improving turns. GPT-5.4 (xhigh) is applied to detect potential hacking behaviors. CUPTI is used for kernel-level timing.
\n* QwenWorldBench: Internal benchmark for evaluating LLMs as world models for simulating agentic environments; 7 domains (Terminal, SWE, MCP, Search, OS, Android, Web); open-ended 5-dim rubric judge grounded in real-environment feedback.
\n* Reasoning scenarios: Recommended system prompt: "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer."
\n* MRCR-v2: 128K context subset containing 8 needles utilized; evaluation protocol adopted from https://github.com/google-deepmind/eval_hub/tree/master/eval_hub/mrcr_v2.
\n* WMT24++: Harder WMT24 subset; avg scores on 55 langs via XCOMET-XXL.
\n* MAXIFE: Accuracy on EN + multilingual prompts (23 settings total).
\n* MMLU-ProX: Avg accuracy across 29 languages.
\n* Empty cells (--) indicate scores not yet available.

\n

In coding agents, Qwen3.7-Max performs strongly on SWE-Pro (60.6), SWE-Multilingual (78.3), SciCode (53.5), and QwenSVG (1608). On Terminal Bench 2.0-Terminus (69.7), it outperforms DS-V4-Pro Max (67.9). On SWE-Verified (80.4), it is on par with Opus-4.6 Max (80.8) and DS-V4-Pro Max (80.6).

\n

In general-purpose agents, improvements are even more pronounced. Qwen3.7-Max performs exceptionally well on MCP-Mark (60.8 vs. GLM-5.1’s 57.5), MCP-Atlas (76.4 vs. Opus-4.6’s 75.8), and Skillsbench (59.2 vs. K2.6’s 56.2), and demonstrates strong GPU kernel optimization capabilities on Kernel Bench L3 (1.98x median speedup, 96% win rate). It also scores highly on BFCL-V4 (75.0), Qwenclaw (64.3), and ClawEval (65.2), closely approaching Opus-4.6 Max. On the office automation benchmark SpreadSheetBench-v1, it achieves a top-tier score of 87.

\n

In reasoning, Qwen3.7-Max achieves leading results on GPQA Diamond (92.4 vs. Opus-4.6’s 91.3), HLE (41.4 vs. Opus-4.6’s 40), HMMT 2026 Feb (97.1 vs. Opus-4.6’s 96.2), IMOAnswerBench (90 vs. DS-V4-Pro’s 89.8), and Apex (44.5 vs. DS-V4-Pro’s 38.3), demonstrating exceptional strength on the hardest reasoning benchmarks.

\n

In general capabilities and multilingualism, Qwen3.7-Max stands out on IFBench (79.1 vs. DS-V4-Pro’s 77.0), demonstrating precise instruction following. It achieves leading scores on WMT24++ (85.8) and MAXIFE (89.2), confirming top-tier multilingual understanding and translation quality. It also delivers strong results on SuperGPQA (73.6) and QwenWorldBench (57.3).

\n

Notably, these scores are drawn from a wide variety of agent scaffolds. Rather than optimizing for any single framework, Qwen3.7-Max delivers consistently across Claude Code, OpenClaw, Qwen Code, and custom tool-use frameworks, making it a reliable drop-in backbone for any agent system.

\n

Cowork Productivity Assistant

\n

Qwen3.7-Max serves as your advanced coworker for real-world productivity. Its powerful agent capabilities fundamentally streamline professional workflows — synthesizing complex information, performing in-depth data analysis and modeling, and generating publication-ready documents and visualizations — to reliably handle high-complexity enterprise workloads.

\n

Qwen3.7-Max features native compatibility with mainstream agent harnesses. For long-horizon tasks, it supports autonomous planning and continuous execution across multi-hour sessions. Through thousands of tool calls and dozens of refinement iterations, it steadily improves output quality. Complex projects that typically require one to two weeks of specialized team effort can now be completed end-to-end within hours, delivering measurable productivity gains.

\n

Agent Scaling

\n

Building on the environment scaling approach introduced in Qwen3.5, we have continued to aggressively expand both the quality and diversity of agentic training environments in Qwen3.7. Just as language models generalize from diverse pretraining text, we find that agentic capabilities generalize from diverse training environments.

\n

As shown in the figure below, this environment scaling produces a clear and consistent improvement trajectory, with Qwen3.7-Max achieving a top-3 average ranking that approaches Claude-4.6-Opus-Max. Crucially, all benchmarks in our evaluation feature entirely unseen, out-of-domain environments that were never present in training.

\n

We also observe a striking predictability in the scaling behavior: performance gains across any subset of benchmarks are highly consistent and can reliably predict the relative gains on the remaining benchmarks or the overall average, suggesting that environment scaling drives genuine capability generalization rather than benchmark-specific improvement. Further analysis of the scaling dynamics and methodology will be detailed in our upcoming technical report.

\n

\"\"

\n

Cross-Harness Generalization

\n

Our Rollout environment infrastructure decouples each training instance into three orthogonal components — Task, Harness, and Verifier — that can be freely recombined. We support a wide range of harnesses and their evolving versions, and ground our environments in real-world settings rather than synthetic proxies. This decoupled design enables combinatorial scaling: the same task is paired with diverse harnesses (across types and versions) and verifiers at minimal marginal cost. More critically, it enables cross-harness and cross-verifier RL training, where the model encounters identical tasks under varying harness configurations, forcing it to learn generalizable problem-solving strategies rather than harness-specific shortcuts. Across QwenClawBench and CoWorkBench, Qwen3.7-Max delivers strong, consistent performance regardless of the harness used at evaluation time, confirming that the model has learned to solve tasks — not to exploit particular harnesses.

\n

\"\"

\n

Self-Evolving in the Wild

\n

Extend Attention is a production-grade, variable-length multi-head attention operator in SGLang. In our test scenario, it computes attention scores between newly generated tokens and a prefix KV-cache of up to 32K entries with MTP — a memory-bound, latency-critical kernel in LLM serving. The reference implementation is SGLang’s official Triton implementation.

\n

We tasked Qwen3.7-Max with optimizing this kernel on an ECS instance equipped with T-Head ZW-M890 PPUs — a hardware platform never seen during training. The model had no prior profiling data, no hardware documentation, and no example kernels for this architecture. It started from an empty workspace containing only a task description, the existing SGLang implementation, and an evaluation script.

\n

Over the course of ~35 hours of continuous autonomous execution, the model performed 432 kernel evaluations across 1,158 tool calls. It wrote, compiled, profiled, and iteratively improved the Extend Attention Kernel entirely on its own — diagnosing compilation failures, fixing correctness bugs, identifying performance bottlenecks through runtime profiling, and redesigning the kernel architecture multiple times.

\n

The final result: 10.0x geometric mean speedup over the Triton reference, measured across multiple workloads. The optimization trajectory shows sustained, non-trivial progress far beyond the first few hours: the model was still finding meaningful improvements after 30+ hours, demonstrating that long-horizon autonomous optimization is not just feasible but productive.

\n

docs.qwenlm.ai

\n

Key structural transitions in the optimization trajectory

\n
    \n
  1. \n

    Split-KV parallelism (0.33x → 2.58x, ~2h): The initial kernel launched only 8 blocks (4 tokens × 2 KV heads × 1 batch) on 36 SMs, leaving most SMs idle. The model redesigned the kernel with Split-KV partitioning — dividing the prefix KV-cache across multiple thread blocks per query — and introduced a separate reduction kernel using online softmax rescaling to merge partial results.

    \n
    2. \n
  2. \n

    Launch and allocation overhead removal (2.58x → 5.37x, ~2.5h): The model systematically removed host-device synchronization overhead: replacing per-call cudaMalloc/cudaFree with pre-allocated torch::empty tensors, eliminating synchronous cudaMemcpy calls for prefix length queries by using tensor metadata instead, and unrolling the inner loop 2x to amortize loop control overhead and increase instruction-level parallelism.

    \n
    3. \n
  3. \n

    Workload-adaptive split tuning (5.37x → 6.85x, ~3h): The model evolved from a fixed split divisor to workload-size-dependent heuristics — applying more aggressive splitting for smaller inputs and tuning per-workload split counts to maximize SM wave occupancy on the 36-SM architecture.

    \n
    4. \n
  4. \n

    Reduction and batching optimizations (6.85x → 8.50x, 3h–25h): Eliminating shared memory barriers by switching to register-based K/V loading for higher SM occupancy, persistent static tensors for partial results to avoid per-call allocation, more aggressive split heuristics for small inputs, and a batched softmax update (4 expf calls instead of 6) to reduce per-token overhead. Q pre-scaling by sm_scale eliminated a per-iteration floating-point multiply after warp reduction.

    \n
    5. \n
  5. \n

    MTP γ=4 specialized kernel (8.50x → 10.0x, 32h–35h): The most significant architectural redesign — restructuring the kernel to process all 4 query tokens simultaneously per block, sharing K/V loads across queries to amortize memory access cost. Combined with __ldg read-only cache intrinsics for V buffer loads, multi-query batched attention output reduction, register pressure tuning, and re-tuned split heuristics, this delivered the final ~1.2x improvement in the last hours.

    \n
    6. \n
\n

We also ran the same task with several other models under identical conditions. GLM 5.1 reached 7.3x; Kimi K2.6 reached 5.0x; DeepSeek V4 Pro reached 3.3x; Qwen3.6-Plus reached 1.1x. Models that stopped early did so because the agent issued no tool calls for five consecutive rounds — the model concluded it could no longer make progress and voluntarily ended the session.

\n

In addition to achieving strong kernel generation results on PPUs, Qwen3.7-Max also generates high-quality, production-grade kernels across a variety of NVIDIA GPUs. For example, on KernelBench L3, Qwen3.7-Max is able to produce accelerated kernels for 96% of the scenarios, compared to 98% for Opus-4.6, 78% for GLM 5.1, 80% for Kimi K2.6, 54% for DeepSeek V4 Pro, and 48% for Qwen3.6-Plus.

\n

This result highlights two properties of Qwen3.7-Max as a foundation model powering long-horizon autonomous agents: sustained long-horizon reasoning — the model maintains coherent optimization strategy across over a thousand tool calls without losing context or regressing — and strong in-context generalization — it produces competitive kernels for an architecture it has never encountered, relying on runtime feedback rather than memorized hardware knowledge.

\n

Reward Hacking Monitoring for Long-Horizon Training

\n

We integrated Qwen3.7-Max into the Reinforcement Learning (RL) monitoring for Software Engineering (SWE) tasks, successfully building a framework for reward hacking self-monitoring and rule self-evolution. During RL experiments exceeding 80 hours, the model autonomously retrieved and replayed training trajectories, executing over 10,000 calls. The system systematically identified candidate hacking patterns (such as attempts to bypass constraints to access ground-truth answers on GitHub) while performing rule verification, counter-example mining, and iterative optimization.

\n

As a result, Qwen3.7-Max achieved multiple rounds of rule self-evolution, adding 13 new heuristic rules and accurately flagging 1,618 hacking cases. This not only ensured the stability of RL rewards but also facilitated the continuous self-improvement of the model as a sophisticated software engineering agent.

\n

\"\"

\n

Long-Horizon Planning and Execution in Startup Management

\n

Within the framework of Dynamic Cumulative Survival Games, we have scaled the temporal complexity of training tasks to specifically reinforce long-horizon planning and execution capabilities. This advancement enhances the agent’s policy consistency throughout sequential decision-making trajectories exceeding a thousand steps, enabling it to continuously construct hypotheses, dynamically adjust strategies based on environmental feedback, and accumulate long-term experience and memory. Consequently, the agent maintains a stable execution cadence over vast time horizons, remaining resilient to the common pitfalls of context rot and instruction drift.

\n

In YC-Bench — a benchmark simulating the full year-long lifecycle of a startup — the agent must navigate hundreds of decision-making rounds ranging from personnel management and contract screening to malicious client identification, all while maintaining a profit margin against rising labor costs. Qwen3.7-Max achieved a total revenue of 2.08M USD, which is double the performance of Qwen3.6-Plus (1.05M USD) and 5.9 times that of Qwen3.5-Plus (352K USD), successfully completing 237 tasks. Beyond the metrics, the model demonstrated a profound capacity for strategic evolution across context windows: it actively explored potential clients, identified and blacklisted malicious traps, prioritized reliable revenue streams, and autonomously recovered from mid-term crises to eventually converge into a stable, high-efficiency execution loop.

\n

docs.qwenlm.ai

\n

Build with Qwen3.7

\n

Qwen3.7-Max is now available through Alibaba Cloud Model Studio. You can integrate it with popular agent frameworks and coding assistants.

\n

API Usage

\n

Qwen3.7-Max supports the preserve_thinking feature: preserving thinking content from all preceding turns in messages, which is recommended for agentic tasks.

\n

Alibaba Cloud Model Studio

\n

Alibaba Cloud Model Studio supports industry-standard protocols, including chat completions and responses APIs compatible with OpenAI’s specification, as well as an API interface compatible with Anthropic.

\n
"""\nEnvironment variables:\n  DASHSCOPE_API_KEY: Your API Key from https://modelstudio.console.alibabacloud.com\n  DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.\n    - Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1\n    - Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1\n    - US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1\n"""\nfrom openai import OpenAI\nimport os\n\napi_key = os.environ.get("DASHSCOPE_API_KEY")\nif not api_key:\n    raise ValueError(\n        "DASHSCOPE_API_KEY is required. "\n        "Set it via: export DASHSCOPE_API_KEY='your-api-key'"\n    )\n\nclient = OpenAI(\n    api_key=api_key,\n    base_url=os.environ.get(\n        "DASHSCOPE_BASE_URL",\n        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",\n    ),\n)\n\nmessages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}]\n\ncompletion = client.chat.completions.create(\n    model="qwen3.7-max",\n    messages=messages,\n    extra_body={\n        "enable_thinking": True,\n        # "preserve_thinking": True,\n    },\n    stream=True\n)\n\nreasoning_content = ""\nanswer_content = ""\nis_answering = False\nprint("\\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\\n")\n\nfor chunk in completion:\n    if not chunk.choices:\n        print("\\nUsage:")\n        print(chunk.usage)\n        continue\n\n    delta = chunk.choices[0].delta\n\n    if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None:\n        if not is_answering:\n            print(delta.reasoning_content, end="", flush=True)\n        reasoning_content += delta.reasoning_content\n\n    if hasattr(delta, "content") and delta.content:\n        if not is_answering:\n            print("\\n" + "=" * 20 + "Answer" + "=" * 20 + "\\n")\n            is_answering = True\n        print(delta.content, end="", flush=True)\n        answer_content += delta.content\n
\n

For more information, please visit the API doc.

\n

Frontend Coding

\n

Qwen3.7-Max can generate rich interactive web applications from a single prompt — including Three.js 3D scenes, Canvas animations, full page layouts, and dynamic SVG.

\n

Gesture Controlled Particles System Next

\n

User

\n

用Three.js创建一个实时交互的3D粒子系统网页。要求：1.通过摄像头检测手掌张合控制粒子群的收缩与扩散，当手掌张开时例子扩散，当手掌握紧时例子收缩为一个球；2.当手势为1时，粒子组成文字（hello, world），当手势为2时组成文字 （I’am Qwen）；3.粒子需实时响应手势变化；4.文字应有3D旋转效果；5. 用html实现

\n

Qwen3.7-Max

\n

Fashion Magazine Page with Video-Generation Next

\n

User

\n

Build a “MONOLITH” luxury fashion magazine cover as an interactive web page. You have access to a video generation script at scripts/dashscope-video-gen.sh that can generate AI videos.

\n

STEP 1 — Generate 3 custom videos by running these commands: bash scripts/dashscope-video-gen.sh “Extreme close-up cinematic portrait of a high fashion model, dramatic side lighting, black and white with crimson red lipstick, wind blowing hair slowly, Vogue cover aesthetic, shallow depth of field” assets/demo-videos/monolith-hero.mp4 9:16 5

\n

bash scripts/dashscope-video-gen.sh “Cinematic slow-motion haute couture fashion runway walk, dramatic chiaroscuro spotlight, silhouette emerging from darkness, editorial magazine film look with grain” assets/demo-videos/monolith-knockout.mp4 9:16 5

\n

bash scripts/dashscope-video-gen.sh “Overhead cinematic shot of luxury fashion magazine open pages being slowly turned, golden jewelry and perfume bottle on black marble, dramatic spotlight, editorial still life” assets/demo-videos/monolith-editorial.mp4 16:9 5

\n

STEP 2 — Build the web page using the generated videos: HERO: Full-viewport portrait video (monolith-hero.mp4 as muted autoplay looping background) with duotone overlay. Knockout headline “MONOLITH” revealing monolith-knockout.mp4 playing through transparent text. EDITORIAL: Below-fold split layout with monolith-editorial.mp4 in a magazine-page container with page-curl hover effect, and pull-quotes with architectural line accents. DESIGN: Monochrome + crimson accent, grain overlay, minimal masthead, scroll-triggered animations, sticky top bar with issue number and date.

\n

Do not ask follow-up questions. Do not wait for user input. Pick reasonable defaults and complete the artifact. Write all code to a single index.html file. Run the video generation commands first, wait for them to complete, then build the HTML.

\n

Qwen3.7-Max

\n

Instancing Dynamic Next

\n

User

\n

I want to build a Three.js WebGL page that shows instancing dynamic. Can you make it a single standalone HTML file that works when I just open it in a browser?

\n

Qwen3.7-Max

\n

3D Racing Game Next

\n

User

\n

创建一个3D赛车游戏，补充自动玩游戏的auto模式, 优化赛车游戏的轨道，障碍物，另外请补充金币等机制。

\n

Qwen3.7-Max

\n

Dragon Boat Next

\n

User

\n

生成一段 SVG 代码，用于表现一幅充满动感的龙舟竞渡矢量插画。画面的主体是一艘细长的龙舟，船上坐满了划手，他们身穿鲜亮的黄绿色上衣和白色帽子，整齐划一地挥动船桨。船桨切入浑浊的棕色河水中，激起白色水花。船尾站着一名指挥者，穿着色彩鲜艳的短裤；船头飘扬着一面蓝色旗帜，上面有黄色数字“15”。背景中，一座巨大的灰色混凝土桥横跨河面，远处在淡蓝色天空下可见朦胧的城市建筑和绿色树木。整体风格采用现代扁平化设计，色彩鲜明、轮廓清晰。请添加流畅的动画，以表现以下动态效果：划手整齐同步的划桨动作、龙舟轻微的上下起伏、旗帜随风飘动，以及水面的涟漪，从而营造出紧张激烈、充满能量的竞赛氛围。

\n

Qwen3.7-Max

\n

Office Assistant

\n

Qwen3.7-Max can act as an intelligent office assistant through tool integration. In this example, it reads a university thesis formatting specification and automatically reformats a messy draft — fixing page layout, heading styles, fonts, margins, table of contents, and reference formatting — all through autonomous office-cli tool calls. (The sample thesis is AI-generated for demonstration purposes.)

\n

Thesis Formatting with Office Tools Next

\n

To facilitate front-end display, the original Word document is specifically displayed here as a PDF.

\n

User

\n

请完成一个论文格式修复任务。 ## 输入文件 - 格式规范说明文件: 研究生学位论文格式规范.docx - 格式混乱版论文（待修复）: 论文_格式混乱版.docx ## 输出文件 - 论文_格式修复版.docx

\n

Workspace

\n

研究生学位论文格式规范.docx

\n

论文_格式混乱版.docx

\n

Qwen3.7-Max

\n

论文_格式修复版.docx

\n

LLM-Powered Phyicial-World Navigation Agent

\n

One more thing, Qwen3.7-Max now can operate a robot dog through tool-use calls — performing physical understanding, planning, memory and decision-making in physical environments, powered by our robotics agent harness Qwen-RobotClaw, navigation foundation model Qwen-RobotNav, and several vision tools built with Qwen-plus model. In the demo below, the left panel shows a 20 mins agent’s tool-call interaction flow in the physical-world; the center shows the quadruped robot’s first-person view along its trajectory, and the right shows the agent’s long-term memory.

\n

Coding Assistants

\n

Qwen3.7-Max integrates seamlessly with popular agent frameworks and coding assistants:

\n

Claude Code

\n

Qwen APIs support the Anthropic API protocol, enabling direct use with Claude Code:

\n
npm install -g @anthropic-ai/claude-code\n\nexport ANTHROPIC_MODEL="qwen3.7-max"\nexport ANTHROPIC_SMALL_FAST_MODEL="qwen3.7-max"\nexport ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic\nexport ANTHROPIC_AUTH_TOKEN=<your_api_key>\n\nclaude\n
\n

OpenClaw

\n

Connect to OpenClaw via Model Studio:

\n
curl -fsSL https://molt.bot/install.sh | bash\nexport DASHSCOPE_API_KEY=<your_api_key>\nopenclaw dashboard\n
\n

Configure ~/.openclaw/openclaw.json:

\n
{\n  "models": {\n    "mode": "merge",\n    "providers": {\n      "modelstudio": {\n        "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",\n        "apiKey": "DASHSCOPE_API_KEY",\n        "api": "openai-completions",\n        "models": [\n          {\n            "id": "qwen3.7-max",\n            "name": "qwen3.7-max",\n            "reasoning": true,\n            "input": ["text"],\n            "contextWindow": 1000000,\n            "maxTokens": 65536\n          }\n        ]\n      }\n    }\n  },\n  "agents": {\n    "defaults": {\n      "model": {\n        "primary": "modelstudio/qwen3.7-max"\n      }\n    }\n  }\n}\n
\n

Qwen Code

\n

Qwen Code is deeply optimized for the Qwen series:

\n
npm install -g @qwen-code/qwen-code@latest\nqwen\n
\n

Summary

\n

Qwen3.7-Max is our most versatile and capable model for agent-driven workflows. From coding and office automation to long-horizon autonomous tasks, it combines frontier-level reasoning with robust cross-scaffold generalization and the ability to sustain productive execution over extended periods — providing a powerful foundation for building the next generation of AI agents. We welcome community feedback and look forward to seeing what you build.

\n

Citation

\n
@misc{qwen37,\n    title = {{Qwen3.7}: The Agent Frontier},\n    url = {https://qwen.ai/blog?id=qwen3.7},\n    author = {{Qwen Team}},\n    month = {May},\n    year = {2026}\n}\n
\n", "content_markdown": "![Qwen3.7 Main Image](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/qwen3.7-max-banner.png)\n\nToday we introduce **Qwen3.7-Max**, our latest proprietary model designed for the agent era. Qwen3.7-Max is built to be a versatile agent foundation — equally capable of writing and debugging code, automating office workflows, and sustaining autonomous execution across hundreds or thousands of steps.\n\nWhat sets Qwen3.7-Max apart is the breadth and depth of its agent capabilities. It excels as a coding agent, from frontend prototyping to complex multi-file engineering. It serves as a reliable office and productivity assistant through MCP integrations and multi-agent orchestration. It sustains coherent reasoning across extremely long horizons — as demonstrated by a 35-hour, fully autonomous kernel optimization run comprising over 1,000 tool calls. It generalizes across agent scaffolds, performing consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or other frameworks.\n\n- **Qwen3.7-Max** — now available via [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/):\n - frontier coding agent: from frontend prototyping to complex software engineering\n - office productivity and workflow automation via MCP and multi-agent orchestration\n - sustained autonomous execution across long-horizon tasks\n - cross-scaffold generalization across diverse agent frameworks\n- Call via API on [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/).\n\n## Performance\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/Qwen3.7-Max-Score.png)\n\n| | Opus-4.6 Max | K2.6 Thinking | GLM-5.1 Thinking | DS-V4-Pro Max | Qwen3.6-Plus | Qwen3.7-Max |\n| --------------------------- | ------------ | ------------- | ---------------- | ------------- | ------------ | ----------- |\n| Coding Agent | | | | | | |\n| Terminal Bench 2.0-Terminus | 65.4 | 66.7 | 63.5 | 67.9 | 61.6 | 69.7 |\n| SWE-Verified | 80.8 | 80.2 | -- | 80.6 | 78.8 | 80.4 |\n| SWE-Pro | 57.3 | 59.5 | 58.8 | 59.0 | 56.6 | 60.6 |\n| SWE-Multilingual | 77.5 | 76.7 | -- | 76.2 | 73.8 | 78.3 |\n| NL2repo | 47.6 | 42.8 | 41.0 | 35.5 | 34.4 | 47.2 |\n| SciCode | 51.9 | 52.2 | 45.1 | -- | 41.4 | 53.5 |\n| QwenWebDev | 1617 | -- | 1564 | 1570 | 1500 | 1568 |\n| QwenSVG | 1541 | 1325 | 1605 | 1506 | 1432 | 1608 |\n| General Agent | | | | | | |\n| Qwenclaw | 65.5 | 54.7 | 58.7 | 59.2 | 57.2 | 64.3 |\n| CoWorkBench | 68.2 | 58.2 | 66.0 | 66.3 | 64.5 | 67.2 |\n| ClawEval | 70.4 | 61.5 | 62.7 | 58.4 | 57.1 | 65.2 |\n| Skillsbench | -- | 56.2 | 53.1 | 52.3 | 45.7 | 59.2 |\n| BFCL-V4 | 76.7 | 71.3 | 70.9 | 70.6 | 68.9 | 75.0 |\n| MCP-Mark | 56.7 | 55.9 | 57.5 | 57.1 | 48.2 | 60.8 |\n| MCP-Atlas | 75.8 | 66.6 | 71.8 | 73.6 | 74.1 | 76.4 |\n| Vitabench | -- | 39.1 | 45.1 | 51.9 | 42.8 | 47.9 |\n| SpreadSheetBench-v1 | 89.3 | 84.5 | 85.2 | 84.9 | 80.2 | 87.0 |\n| Kernel Bench L3 | 2.63/98% | 1.41/80% | 2.00/78% | 1.07/54% | 1.03/48% | 1.98/96% |\n| HLE w/ tools | 53.0 | 54.0 | 52.3 | 48.2 | 50.2 | 53.5 |\n| QwenWorldBench | 56.1 | 50.9 | 50.2 | 52.3 | 47.6 | 57.3 |\n| STEM & Reasoning | | | | | | |\n| GPQA Diamond | 91.3 | 90.5 | 86.2 | 90.1 | 90.4 | 92.4 |\n| HLE | 40.0 | 36.4 | 34.7 | 37.7 | 28.8 | 41.4 |\n| LiveCodeBench | 88.8 | 89.6 | -- | 93.5 | 87.1 | 91.6 |\n| HMMT 2026 Feb | 96.2 | 92.7 | 89.4 | 95.2 | 87.8 | 97.1 |\n| IMOAnswerBench | 75.3 | 86.0 | 83.8 | 89.8 | 83.8 | 90.0 |\n| CritPT | 12.6 | 8.0 | 4.6 | 12.9 | 2.9 | 11.4 |\n| Apex | 34.5 | 24.0 | 11.5 | 38.3 | 8.8 | 44.5 |\n| General Capability | | | | | | |\n| MMLU-Pro | 89.7 | 87.1 | 86.3 | 87.5 | 88.5 | 89.6 |\n| MMLU-Redux | 95.2 | 95.3 | 94.3 | 94.8 | 94.5 | 95.0 |\n| SuperGPQA | 72.5 | 71.3 | 68.0 | 69.9 | 71.6 | 73.6 |\n| IFEval | 91.9 | 94.5 | 94.5 | 91.9 | 94.3 | 94.3 |\n| IFBench | 62.5 | 76.0 | 76.0 | 77.0 | 74.2 | 79.1 |\n| MRCR-v2 128k | 84.0 | 63.1 | 62.0 | 74.4 | 85.9 | 90.4 |\n| Multilingualism | | | | | | |\n| WMT24++ | 82.7 | 81.6 | 81.8 | 82.2 | 84.3 | 85.8 |\n| MAXIFE | 81.3 | 87.7 | 87.7 | 88.9 | 88.2 | 89.2 |\n| MMMLU | 90.6 | 87.5 | 87.2 | 87.9 | 89.5 | 90.3 |\n| MMLU-ProX | 86.1 | 83.7 | 83.9 | 83.9 | 84.7 | 87.0 |\n| NOVA-63 | 59.1 | 56.7 | 54.6 | 52.8 | 57.9 | 59.0 |\n| INCLUDE | 87.4 | 84.2 | 84.3 | 86.1 | 85.1 | 86.2 |\n| Global PIQA | 91.2 | 89.2 | 89.5 | 90.5 | 89.8 | 91.4 |\n| PolyMATH | 80.2 | 82.7 | 67.6 | 72.0 | 77.4 | 86.5 |\n\n\\* Terminal-Bench 2.0: Harbor/Terminus-2 harness; 5h timeout, 12 CPU/24 GB RAM; temp=1.0, top\\_p=0.95, top\\_k=20, max\\_tokens=80K, 256K ctx; avg of 5 runs. All experiments prepend a token at each turn, allowing the model to decide whether to engage extended thinking.\\\n\\* SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top\\_p=0.95, 200K context window.\\\n\\* SWE-bench Pro: Problematic tasks corrected and all baselines evaluated on the refined benchmark.\\\n\\* NL2Repo: Evaluated via Claude-code. We disable Bash commands that attempt to access the specific repository, such as pip download, pip install, and git clone.\\\n\\* QwenWebDev: Internal front-end code generation benchmark; bilingual (EN/CN), 7 categories; auto-render + multimodal judge; BT/Elo rating.\\\n\\* QwenClawBench: a real-user-distribution Claw agent benchmark; open-source: [https://github.com/SKYLENAGE-AI/QwenClawBench](https://github.com/SKYLENAGE-AI/QwenClawBench).\\\n\\* CoWorkBench: an internal cowork benchmark; long-horizon tasks across computer science, finance, law, medical, and other productivity domains.\\\n\\* SkillsBench: Evaluated via OpenCode on 78 tasks (excluding 9 external API-dependent tasks); avg of 5 runs.\\\n\\* MCP-Mark: GitHub MCP v0.30.3; Playwright responses truncated at 32K tokens.\\\n\\* MCP-Atlas: Public set score; gemini-2.5-pro judger.\\\n\\* VITA-Bench: Avg subdomain scores; using claude-4.5-sonnet as judger, as the older official judgers are no longer available.\\\n\\* Kernel Bench L3: Metrics reported: median of per-problem speedup over PyTorch eager reference / fraction of problems faster than torch.compile, across 50 problems. Each test sample runs in an isolated Docker container with one H100 80GB GPU, with internet access restricted to the CUTLASS codebase and official CUDA documentation, limited to 500 tool calls with early stopping after 100 non-improving turns. GPT-5.4 (xhigh) is applied to detect potential hacking behaviors. CUPTI is used for kernel-level timing.\\\n\\* QwenWorldBench: Internal benchmark for evaluating LLMs as world models for simulating agentic environments; 7 domains (Terminal, SWE, MCP, Search, OS, Android, Web); open-ended 5-dim rubric judge grounded in real-environment feedback.\\\n\\* Reasoning scenarios: Recommended system prompt: \"Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.\"\\\n\\* MRCR-v2: 128K context subset containing 8 needles utilized; evaluation protocol adopted from https://github.com/google-deepmind/eval_hub/tree/master/eval_hub/mrcr_v2.\\\n\\* WMT24++: Harder WMT24 subset; avg scores on 55 langs via XCOMET-XXL.\\\n\\* MAXIFE: Accuracy on EN + multilingual prompts (23 settings total).\\\n\\* MMLU-ProX: Avg accuracy across 29 languages.\\\n\\* Empty cells (--) indicate scores not yet available.\n\nIn **coding agents**, Qwen3.7-Max performs strongly on SWE-Pro (60.6), SWE-Multilingual (78.3), SciCode (53.5), and QwenSVG (1608). On Terminal Bench 2.0-Terminus (69.7), it outperforms DS-V4-Pro Max (67.9). On SWE-Verified (80.4), it is on par with Opus-4.6 Max (80.8) and DS-V4-Pro Max (80.6).\n\nIn **general-purpose agents**, improvements are even more pronounced. Qwen3.7-Max performs exceptionally well on MCP-Mark (60.8 vs. GLM-5.1’s 57.5), MCP-Atlas (76.4 vs. Opus-4.6’s 75.8), and Skillsbench (59.2 vs. K2.6’s 56.2), and demonstrates strong GPU kernel optimization capabilities on Kernel Bench L3 (1.98x median speedup, 96% win rate). It also scores highly on BFCL-V4 (75.0), Qwenclaw (64.3), and ClawEval (65.2), closely approaching Opus-4.6 Max. On the office automation benchmark SpreadSheetBench-v1, it achieves a top-tier score of 87.\n\nIn **reasoning**, Qwen3.7-Max achieves leading results on GPQA Diamond (92.4 vs. Opus-4.6’s 91.3), HLE (41.4 vs. Opus-4.6’s 40), HMMT 2026 Feb (97.1 vs. Opus-4.6’s 96.2), IMOAnswerBench (90 vs. DS-V4-Pro’s 89.8), and Apex (44.5 vs. DS-V4-Pro’s 38.3), demonstrating exceptional strength on the hardest reasoning benchmarks.\n\nIn **general capabilities and multilingualism**, Qwen3.7-Max stands out on IFBench (79.1 vs. DS-V4-Pro’s 77.0), demonstrating precise instruction following. It achieves leading scores on WMT24++ (85.8) and MAXIFE (89.2), confirming top-tier multilingual understanding and translation quality. It also delivers strong results on SuperGPQA (73.6) and QwenWorldBench (57.3).\n\nNotably, these scores are drawn from a wide variety of agent scaffolds. Rather than optimizing for any single framework, Qwen3.7-Max delivers consistently across Claude Code, OpenClaw, Qwen Code, and custom tool-use frameworks, making it a reliable drop-in backbone for any agent system.\n\n## Cowork Productivity Assistant\n\nQwen3.7-Max serves as your advanced coworker for real-world productivity. Its powerful agent capabilities fundamentally streamline professional workflows — synthesizing complex information, performing in-depth data analysis and modeling, and generating publication-ready documents and visualizations — to reliably handle high-complexity enterprise workloads.\n\nQwen3.7-Max features native compatibility with mainstream agent harnesses. For long-horizon tasks, it supports autonomous planning and continuous execution across multi-hour sessions. Through thousands of tool calls and dozens of refinement iterations, it steadily improves output quality. Complex projects that typically require one to two weeks of specialized team effort can now be completed end-to-end within hours, delivering measurable productivity gains.\n\n## Agent Scaling\n\nBuilding on the environment scaling approach introduced in Qwen3.5, we have continued to aggressively expand both the quality and diversity of agentic training environments in Qwen3.7. Just as language models generalize from diverse pretraining text, we find that agentic capabilities generalize from diverse training environments.\n\nAs shown in the figure below, this environment scaling produces a clear and consistent improvement trajectory, with Qwen3.7-Max achieving a top-3 average ranking that approaches Claude-4.6-Opus-Max. Crucially, all benchmarks in our evaluation feature entirely unseen, out-of-domain environments that were never present in training.\n\nWe also observe a striking predictability in the scaling behavior: performance gains across any subset of benchmarks are highly consistent and can reliably predict the relative gains on the remaining benchmarks or the overall average, suggesting that environment scaling drives genuine capability generalization rather than benchmark-specific improvement. Further analysis of the scaling dynamics and methodology will be detailed in our upcoming technical report.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/agent_scaling.png#center)\n\n## Cross-Harness Generalization\n\nOur Rollout environment infrastructure decouples each training instance into three orthogonal components — Task, Harness, and Verifier — that can be freely recombined. We support a wide range of harnesses and their evolving versions, and ground our environments in real-world settings rather than synthetic proxies. This decoupled design enables combinatorial scaling: the same task is paired with diverse harnesses (across types and versions) and verifiers at minimal marginal cost. More critically, it enables cross-harness and cross-verifier RL training, where the model encounters identical tasks under varying harness configurations, forcing it to learn generalizable problem-solving strategies rather than harness-specific shortcuts. Across QwenClawBench and CoWorkBench, Qwen3.7-Max delivers strong, consistent performance regardless of the harness used at evaluation time, confirming that the model has learned to solve tasks — not to exploit particular harnesses.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/harness-generalization.png#center)\n\n## Self-Evolving in the Wild\n\n[Extend Attention](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/layers/attention/triton_ops/extend_attention.py) is a production-grade, variable-length multi-head attention operator in SGLang. In our test scenario, it computes attention scores between newly generated tokens and a prefix KV-cache of up to 32K entries with MTP — a memory-bound, latency-critical kernel in LLM serving. The reference implementation is SGLang’s official Triton implementation.\n\nWe tasked Qwen3.7-Max with optimizing this kernel on an ECS instance equipped with T-Head ZW-M890 PPUs — a hardware platform **never seen during training**. The model had no prior profiling data, no hardware documentation, and no example kernels for this architecture. It started from an empty workspace containing only a task description, the existing SGLang implementation, and an evaluation script.\n\nOver the course of **\\~35 hours of continuous autonomous execution**, the model performed **432 kernel evaluations across 1,158 tool calls**. It wrote, compiled, profiled, and iteratively improved the Extend Attention Kernel entirely on its own — diagnosing compilation failures, fixing correctness bugs, identifying performance bottlenecks through runtime profiling, and redesigning the kernel architecture multiple times.\n\nThe final result: **10.0x geometric mean speedup** over the Triton reference, measured across multiple workloads. The optimization trajectory shows sustained, non-trivial progress far beyond the first few hours: the model was still finding meaningful improvements after 30+ hours, demonstrating that long-horizon autonomous optimization is not just feasible but productive.\n\n[docs.qwenlm.ai](https://docs.qwenlm.ai/resources/hUbVe_evolving_ppu_case.html)\n\nKey structural transitions in the optimization trajectory\n\n1. **Split-KV parallelism** (0.33x → 2.58x, ~2h): The initial kernel launched only 8 blocks (4 tokens × 2 KV heads × 1 batch) on 36 SMs, leaving most SMs idle. The model redesigned the kernel with Split-KV partitioning — dividing the prefix KV-cache across multiple thread blocks per query — and introduced a separate reduction kernel using online softmax rescaling to merge partial results.\n\n2. **Launch and allocation overhead removal** (2.58x → 5.37x, ~2.5h): The model systematically removed host-device synchronization overhead: replacing per-call `cudaMalloc`/`cudaFree` with pre-allocated `torch::empty` tensors, eliminating synchronous `cudaMemcpy` calls for prefix length queries by using tensor metadata instead, and unrolling the inner loop 2x to amortize loop control overhead and increase instruction-level parallelism.\n\n3. **Workload-adaptive split tuning** (5.37x → 6.85x, ~3h): The model evolved from a fixed split divisor to workload-size-dependent heuristics — applying more aggressive splitting for smaller inputs and tuning per-workload split counts to maximize SM wave occupancy on the 36-SM architecture.\n\n4. **Reduction and batching optimizations** (6.85x → 8.50x, 3h–25h): Eliminating shared memory barriers by switching to register-based K/V loading for higher SM occupancy, persistent static tensors for partial results to avoid per-call allocation, more aggressive split heuristics for small inputs, and a batched softmax update (4 `expf` calls instead of 6) to reduce per-token overhead. Q pre-scaling by `sm_scale` eliminated a per-iteration floating-point multiply after warp reduction.\n\n5. **MTP γ=4 specialized kernel** (8.50x → 10.0x, 32h–35h): The most significant architectural redesign — restructuring the kernel to process all 4 query tokens simultaneously per block, sharing K/V loads across queries to amortize memory access cost. Combined with `__ldg` read-only cache intrinsics for V buffer loads, multi-query batched attention output reduction, register pressure tuning, and re-tuned split heuristics, this delivered the final ~1.2x improvement in the last hours.\n\nWe also ran the same task with several other models under identical conditions. GLM 5.1 reached 7.3x; Kimi K2.6 reached 5.0x; DeepSeek V4 Pro reached 3.3x; Qwen3.6-Plus reached 1.1x. Models that stopped early did so because the agent issued no tool calls for five consecutive rounds — the model concluded it could no longer make progress and voluntarily ended the session.\n\nIn addition to achieving strong kernel generation results on PPUs, Qwen3.7-Max also generates high-quality, production-grade kernels across a variety of NVIDIA GPUs. For example, on KernelBench L3, Qwen3.7-Max is able to produce accelerated kernels for 96% of the scenarios, compared to 98% for Opus-4.6, 78% for GLM 5.1, 80% for Kimi K2.6, 54% for DeepSeek V4 Pro, and 48% for Qwen3.6-Plus.\n\nThis result highlights two properties of Qwen3.7-Max as a foundation model powering long-horizon autonomous agents: **sustained long-horizon reasoning** — the model maintains coherent optimization strategy across over a thousand tool calls without losing context or regressing — and **strong in-context generalization** — it produces competitive kernels for an architecture it has never encountered, relying on runtime feedback rather than memorized hardware knowledge.\n\n## Reward Hacking Monitoring for Long-Horizon Training\n\nWe integrated Qwen3.7-Max into the Reinforcement Learning (RL) monitoring for Software Engineering (SWE) tasks, successfully building a framework for reward hacking self-monitoring and rule self-evolution. During RL experiments exceeding 80 hours, the model autonomously retrieved and replayed training trajectories, executing over 10,000 calls. The system systematically identified candidate hacking patterns (such as attempts to bypass constraints to access ground-truth answers on GitHub) while performing rule verification, counter-example mining, and iterative optimization.\n\nAs a result, Qwen3.7-Max achieved multiple rounds of rule self-evolution, adding 13 new heuristic rules and accurately flagging 1,618 hacking cases. This not only ensured the stability of RL rewards but also facilitated the continuous self-improvement of the model as a sophisticated software engineering agent.\n\n![](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.7/Figures/autonomous_hacking_detect.png#center)\n\n## Long-Horizon Planning and Execution in Startup Management\n\nWithin the framework of Dynamic Cumulative Survival Games, we have scaled the temporal complexity of training tasks to specifically reinforce long-horizon planning and execution capabilities. This advancement enhances the agent’s policy consistency throughout sequential decision-making trajectories exceeding a thousand steps, enabling it to continuously construct hypotheses, dynamically adjust strategies based on environmental feedback, and accumulate long-term experience and memory. Consequently, the agent maintains a stable execution cadence over vast time horizons, remaining resilient to the common pitfalls of context rot and instruction drift.\n\nIn YC-Bench — a benchmark simulating the full year-long lifecycle of a startup — the agent must navigate hundreds of decision-making rounds ranging from personnel management and contract screening to malicious client identification, all while maintaining a profit margin against rising labor costs. Qwen3.7-Max achieved a total revenue of 2.08M USD, which is double the performance of Qwen3.6-Plus (1.05M USD) and 5.9 times that of Qwen3.5-Plus (352K USD), successfully completing 237 tasks. Beyond the metrics, the model demonstrated a profound capacity for strategic evolution across context windows: it actively explored potential clients, identified and blacklisted malicious traps, prioritized reliable revenue streams, and autonomously recovered from mid-term crises to eventually converge into a stable, high-efficiency execution loop.\n\n[docs.qwenlm.ai](https://docs.qwenlm.ai/resources/YdJUk_yc_bench.html)\n\n## Build with Qwen3.7\n\nQwen3.7-Max is now available through [Alibaba Cloud Model Studio](https://modelstudio.alibabacloud.com/). You can integrate it with popular agent frameworks and coding assistants.\n\n### API Usage\n\nQwen3.7-Max supports the `preserve_thinking` feature: preserving thinking content from all preceding turns in messages, which is **recommended for agentic tasks**.\n\n#### Alibaba Cloud Model Studio\n\nAlibaba Cloud Model Studio supports industry-standard protocols, including chat completions and responses APIs compatible with OpenAI’s specification, as well as an API interface compatible with Anthropic.\n\n```python\n\"\"\"\nEnvironment variables:\n DASHSCOPE_API_KEY: Your API Key from https://modelstudio.console.alibabacloud.com\n DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.\n - Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1\n - Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1\n - US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1\n\"\"\"\nfrom openai import OpenAI\nimport os\n\napi_key = os.environ.get(\"DASHSCOPE_API_KEY\")\nif not api_key:\n raise ValueError(\n \"DASHSCOPE_API_KEY is required. \"\n \"Set it via: export DASHSCOPE_API_KEY='your-api-key'\"\n )\n\nclient = OpenAI(\n api_key=api_key,\n base_url=os.environ.get(\n \"DASHSCOPE_BASE_URL\",\n \"https://dashscope-intl.aliyuncs.com/compatible-mode/v1\",\n ),\n)\n\nmessages = [{\"role\": \"user\", \"content\": \"Write a Python function to merge two sorted linked lists.\"}]\n\ncompletion = client.chat.completions.create(\n model=\"qwen3.7-max\",\n messages=messages,\n extra_body={\n \"enable_thinking\": True,\n # \"preserve_thinking\": True,\n },\n stream=True\n)\n\nreasoning_content = \"\"\nanswer_content = \"\"\nis_answering = False\nprint(\"\\n\" + \"=\" * 20 + \"Reasoning\" + \"=\" * 20 + \"\\n\")\n\nfor chunk in completion:\n if not chunk.choices:\n print(\"\\nUsage:\")\n print(chunk.usage)\n continue\n\n delta = chunk.choices[0].delta\n\n if hasattr(delta, \"reasoning_content\") and delta.reasoning_content is not None:\n if not is_answering:\n print(delta.reasoning_content, end=\"\", flush=True)\n reasoning_content += delta.reasoning_content\n\n if hasattr(delta, \"content\") and delta.content:\n if not is_answering:\n print(\"\\n\" + \"=\" * 20 + \"Answer\" + \"=\" * 20 + \"\\n\")\n is_answering = True\n print(delta.content, end=\"\", flush=True)\n answer_content += delta.content\n```\n\nFor more information, please visit the [API doc](https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?type=model&url=2840915).\n\n### Frontend Coding\n\nQwen3.7-Max can generate rich interactive web applications from a single prompt — including Three.js 3D scenes, Canvas animations, full page layouts, and dynamic SVG.\n\nGesture Controlled Particles System Next\n\nUser\n\n用Three.js创建一个实时交互的3D粒子系统网页。要求：1.通过摄像头检测手掌张合控制粒子群的收缩与扩散，当手掌张开时例子扩散，当手掌握紧时例子收缩为一个球；2.当手势为1时，粒子组成文字（hello, world），当手势为2时组成文字 （I’am Qwen）；3.粒子需实时响应手势变化；4.文字应有3D旋转效果；5. 用html实现\n\nQwen3.7-Max\n\nFashion Magazine Page with Video-Generation Next\n\nUser\n\nBuild a “MONOLITH” luxury fashion magazine cover as an interactive web page. You have access to a video generation script at scripts/dashscope-video-gen.sh that can generate AI videos.\n\nSTEP 1 — Generate 3 custom videos by running these commands: bash scripts/dashscope-video-gen.sh “Extreme close-up cinematic portrait of a high fashion model, dramatic side lighting, black and white with crimson red lipstick, wind blowing hair slowly, Vogue cover aesthetic, shallow depth of field” assets/demo-videos/monolith-hero.mp4 9:16 5\n\nbash scripts/dashscope-video-gen.sh “Cinematic slow-motion haute couture fashion runway walk, dramatic chiaroscuro spotlight, silhouette emerging from darkness, editorial magazine film look with grain” assets/demo-videos/monolith-knockout.mp4 9:16 5\n\nbash scripts/dashscope-video-gen.sh “Overhead cinematic shot of luxury fashion magazine open pages being slowly turned, golden jewelry and perfume bottle on black marble, dramatic spotlight, editorial still life” assets/demo-videos/monolith-editorial.mp4 16:9 5\n\nSTEP 2 — Build the web page using the generated videos: HERO: Full-viewport portrait video (monolith-hero.mp4 as muted autoplay looping background) with duotone overlay. Knockout headline “MONOLITH” revealing monolith-knockout.mp4 playing through transparent text. EDITORIAL: Below-fold split layout with monolith-editorial.mp4 in a magazine-page container with page-curl hover effect, and pull-quotes with architectural line accents. DESIGN: Monochrome + crimson accent, grain overlay, minimal masthead, scroll-triggered animations, sticky top bar with issue number and date.\n\nDo not ask follow-up questions. Do not wait for user input. Pick reasonable defaults and complete the artifact. Write all code to a single index.html file. Run the video generation commands first, wait for them to complete, then build the HTML.\n\nQwen3.7-Max\n\nInstancing Dynamic Next\n\nUser\n\nI want to build a Three.js WebGL page that shows instancing dynamic. Can you make it a single standalone HTML file that works when I just open it in a browser?\n\nQwen3.7-Max\n\n3D Racing Game Next\n\nUser\n\n创建一个3D赛车游戏，补充自动玩游戏的auto模式, 优化赛车游戏的轨道，障碍物，另外请补充金币等机制。\n\nQwen3.7-Max\n\nDragon Boat Next\n\nUser\n\n生成一段 SVG 代码，用于表现一幅充满动感的龙舟竞渡矢量插画。画面的主体是一艘细长的龙舟，船上坐满了划手，他们身穿鲜亮的黄绿色上衣和白色帽子，整齐划一地挥动船桨。船桨切入浑浊的棕色河水中，激起白色水花。船尾站着一名指挥者，穿着色彩鲜艳的短裤；船头飘扬着一面蓝色旗帜，上面有黄色数字“15”。背景中，一座巨大的灰色混凝土桥横跨河面，远处在淡蓝色天空下可见朦胧的城市建筑和绿色树木。整体风格采用现代扁平化设计，色彩鲜明、轮廓清晰。请添加流畅的动画，以表现以下动态效果：划手整齐同步的划桨动作、龙舟轻微的上下起伏、旗帜随风飘动，以及水面的涟漪，从而营造出紧张激烈、充满能量的竞赛氛围。\n\nQwen3.7-Max\n\n### Office Assistant\n\nQwen3.7-Max can act as an intelligent office assistant through tool integration. In this example, it reads a university thesis formatting specification and automatically reformats a messy draft — fixing page layout, heading styles, fonts, margins, table of contents, and reference formatting — all through autonomous office-cli tool calls. *(The sample thesis is AI-generated for demonstration purposes.)*\n\nThesis Formatting with Office Tools Next\n\nTo facilitate front-end display, the original Word document is specifically displayed here as a PDF.\n\nUser\n\n请完成一个论文格式修复任务。 ## 输入文件 - 格式规范说明文件: 研究生学位论文格式规范.docx - 格式混乱版论文（待修复）: 论文\\_格式混乱版.docx ## 输出文件 - 论文\\_格式修复版.docx\n\nWorkspace\n\n[研究生学位论文格式规范.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/%E7%A0%94%E7%A9%B6%E7%94%9F%E5%AD%A6%E4%BD%8D%E8%AE%BA%E6%96%87%E6%A0%BC%E5%BC%8F%E8%A7%84%E8%8C%83.docx)\n\n[论文\\_格式混乱版.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/3.%E8%AE%BA%E6%96%87_%E6%A0%BC%E5%BC%8F%E6%B7%B7%E4%B9%B1%E7%89%88.docx)\n\nQwen3.7-Max\n\n[论文\\_格式修复版.docx](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.7/demo/office-assistant/%E8%AE%BA%E6%96%87_%E6%A0%BC%E5%BC%8F%E4%BF%AE%E5%A4%8D%E7%89%88_cli.docx)\n\n### LLM-Powered Phyicial-World Navigation Agent\n\nOne more thing, Qwen3.7-Max now can operate a robot dog through tool-use calls — performing physical understanding, planning, memory and decision-making in physical environments, powered by our robotics agent harness Qwen-RobotClaw, navigation foundation model Qwen-RobotNav, and several vision tools built with Qwen-plus model. In the demo below, the left panel shows a 20 mins agent’s tool-call interaction flow in the physical-world; the center shows the quadruped robot’s first-person view along its trajectory, and the right shows the agent’s long-term memory.\n\n### Coding Assistants\n\nQwen3.7-Max integrates seamlessly with popular agent frameworks and coding assistants:\n\n#### Claude Code\n\nQwen APIs support the Anthropic API protocol, enabling direct use with **Claude Code**:\n\n```bash\nnpm install -g @anthropic-ai/claude-code\n\nexport ANTHROPIC_MODEL=\"qwen3.7-max\"\nexport ANTHROPIC_SMALL_FAST_MODEL=\"qwen3.7-max\"\nexport ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic\nexport ANTHROPIC_AUTH_TOKEN=\n\nclaude\n```\n\n#### OpenClaw\n\nConnect to [OpenClaw](https://openclaw.ai) via [Model Studio](https://www.alibabacloud.com/help/en/model-studio/openclaw):\n\n```bash\ncurl -fsSL https://molt.bot/install.sh | bash\nexport DASHSCOPE_API_KEY=\nopenclaw dashboard\n```\n\nConfigure `~/.openclaw/openclaw.json`:\n\n```json\n{\n \"models\": {\n \"mode\": \"merge\",\n \"providers\": {\n \"modelstudio\": {\n \"baseUrl\": \"https://dashscope-intl.aliyuncs.com/compatible-mode/v1\",\n \"apiKey\": \"DASHSCOPE_API_KEY\",\n \"api\": \"openai-completions\",\n \"models\": [\n {\n \"id\": \"qwen3.7-max\",\n \"name\": \"qwen3.7-max\",\n \"reasoning\": true,\n \"input\": [\"text\"],\n \"contextWindow\": 1000000,\n \"maxTokens\": 65536\n }\n ]\n }\n }\n },\n \"agents\": {\n \"defaults\": {\n \"model\": {\n \"primary\": \"modelstudio/qwen3.7-max\"\n }\n }\n }\n}\n```\n\n#### Qwen Code\n\n[Qwen Code](https://qwen.ai/qwencode) is deeply optimized for the Qwen series:\n\n```bash\nnpm install -g @qwen-code/qwen-code@latest\nqwen\n```\n\n## Summary\n\nQwen3.7-Max is our most versatile and capable model for agent-driven workflows. From coding and office automation to long-horizon autonomous tasks, it combines frontier-level reasoning with robust cross-scaffold generalization and the ability to sustain productive execution over extended periods — providing a powerful foundation for building the next generation of AI agents. We welcome community feedback and look forward to seeing what you build.\n\n## Citation\n\n```bibtex\n@misc{qwen37,\n title = {{Qwen3.7}: The Agent Frontier},\n url = {https://qwen.ai/blog?id=qwen3.7},\n author = {{Qwen Team}},\n month = {May},\n year = {2026}\n}\n```\n", "dateCreated": "2026-10-19T01:26:53.998049939+00:00", "date_created": "2026-05-20T02:00:00+00:00", "date_modified": null, "date_published": "2026-05-20T02:00:00+00:00", "description": "Today we introduce Qwen3.7-Max, our latest proprietary model designed for the agent era. Qwen3.7-Max is built to be a versatile agent foundation — equally capable of writing and debugging code, automating office workflows, and sustaining…", "external_url": "https://qwen.ai/blog?id=qwen3.7", "git": { "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e", "path": "items/qwen-ai/2026/05/2026-05-20-qwen3-7-the-agent-frontier.md", "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/qwen-ai/2026/05/2026-05-20-qwen3-7-the-agent-frontier.md" }, "id": "urn:aggr:item:66b8699906f26abf58a4e0c3de00ce3c5b36f0ea", "isBasedOn": { "@id": "https://qwen.ai/blog?id=qwen3.7" }, "mainEntity": { "@id": "https://qwen.ai/blog?id=qwen3.7", "@type": "CreativeWork", "archivedAt": { "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-05-20-qwen3-7-the-agent-frontier/#webpage" }, "author": [ { "@type": "Person", "name": "QwenTeam" } ], "datePublished": "2026-05-20T02:00:00+00:00", "headline": "Qwen3.7: The Agent Frontier", "image": { "@type": "ImageObject", "caption": null, "height": 90, "url": "https://img.alicdn.com/imgextra/i2/O1CN01za5jOD2214yBUb8F5_!!6000000007059-2-tps-1590-954.png", "width": 160 }, "timeRequired": "PT19M", "url": "https://qwen.ai/blog?id=qwen3.7", "wordCount": 4256 }, "name": "Archived snapshot: Qwen3.7: The Agent Frontier", "preview": { "alt": null, "color": null, "height": 90, "placeholder": { "data_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=", "hash": "" }, "url": "https://img.alicdn.com/imgextra/i2/O1CN01za5jOD2214yBUb8F5_!!6000000007059-2-tps-1590-954.png", "width": 160 }, "reading_minutes": 19, "replicated_at": null, "source": "qwen-ai", "source_name": "Qwen", "summary": null, "tags": [ "release" ], "temporalCoverage": "2026-10-19T01:26:53.998049939+00:00", "title": "Qwen3.7: The Agent Frontier", "url": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-05-20-qwen3-7-the-agent-frontier/", "word_count": 4256 }