{
  "@context": "https://schema.org",
  "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-02-02-qwen3-coder-next-pushing-small-hybrid-models-on-agentic/#webpage",
  "@type": [
    "WebPage",
    "ArchiveComponent"
  ],
  "articleBody": "## [Tech Report](https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf) [GitHub](https://github.com/QwenLM/Qwen3-Coder) [Hugging Face](https://huggingface.co/collections/Qwen/qwen3-coder-next) [ModelScope](https://modelscope.cn/collections/Qwen/Qwen3-Coder-Next) [DISCORD](https://discord.gg/yPEP2vHTu4)\n\n## Introduction\n\nWe introduce **Qwen3-Coder-Next**, an open-weight language model designed specifically for coding agents and local development. Built on top of **Qwen3-Next-80B-A3B-Base**, which adopts a novel architecture with hybrid attention and MoE, Qwen3-Coder-Next has been agentically trained at scale on large-scale executable task synthesis, environment interaction, and reinforcement learning, obtaining strong coding and agentic capabilities with significantly lower inference costs.\n\n## Scaling Agentic Training\n\nRather than relying solely on parameter scaling, Qwen3-Coder-Next focuses on **scaling agentic training signals**. We train the model using large collections of *verifiable coding tasks* paired with executable environments, enabling the model to learn directly from environment feedback. This includes:\n\n- Continued pretraining on code- and agent-centric data\n- Supervised fine-tuning on data containing high-quality agent trajectories\n- Domain-specialized expert training (e.g., software engineering, QA, web/UX)\n- Expert distillation into a single deployment-ready model\n\nThis recipe emphasizes **long-horizon reasoning, tool usage, and recovery from execution failures**, which are essential for real-world coding agents.\n\n## Performance on Coding Agent Benchmarks\n\n### Agent-Centric Benchmark Results\n\nThe figure below summarizes performance across several widely used **coding agent benchmarks**, including SWE-Bench (Verified, Multilingual, and Pro), TerminalBench 2.0, and Aider.\n\n![Qwen3 Coder Next Benchmarks](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/benchmarks.png)\n\nThe figure demonstrates that:\n\n- Qwen3-Coder-Next achieves **over 70% on SWE-Bench Verified** using the SWE-Agent scaffold.\n- Performance remains competitive across **multilingual settings** and the more challenging **SWE-Bench Pro** benchmark.\n- Despite its small active footprint, the model matches or exceeds several much larger open-source models across agent-centric evaluations.\n\nAs shown in the figure below, Our model achieves strong results on SWE-Bench Pro by scaling the number of agent turns, providing evidence that the model excels at long-horizon reasoning in multi-turn agentic tasks.\n\n![SWE-Bench Pro turns](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro_vs_turns.png)\n\n### Efficiency–Performance Tradeoff\n\nThis figure highlights how Qwen3-Coder-Next achieves an improved Pareto tradeoff between efficiency and performance.\n\n![Qwen3 Coder Next Main Image](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro.png)\n\nThis comparison makes the efficiency story clear:\n\n- **Qwen3-Coder-Next (3B active)** achieves SWE-Bench-Pro performance comparable to models with **10×–20× more active parameters**.\n- Qwen3-Coder-Next sits on a strong Pareto frontier for **cost-effective agent deployment**.\n\n## Demo\n\nThe small and fast coder model can be integrated into different downstream applications, and below we demonstrate showcases of OpenClaw, Qwen Code, Claude Code, Web dev, browser use, Cline, etc.\n\n### Web Dev\n\nCreating a Chat Interface Next\n\nCreate an Interactive Game Next\n\nBuilding an Interactive ASCII Art Drawing Tool Next\n\n### CLI\n\nDesktop Cleanup Next\n\nImplementing a Web Game: Zombies vs. Plants Next\n\n### Cline\n\nCreating a Multicolor Animation Next\n\nBuilding an Interactive Sound Art Tool Next\n\n### OpenClaw\n\nWeb Page for Qwen3 Coder Next Next\n\n### Browser Use Agent\n\nSearching for a Product on Amazon Next\n\nTesting a Website Next\n\n### coder.qwen.ai\n\n## Summary and Future Work\n\nQwen3-Coder-Next shows promising results on coding agent benchmarks, providing good speed and reasoning abilities for practical use. While it performs competitively—even compared to some larger open-source models—there is still much room for improvement.\n\nLooking ahead, we believe strong agent skills—like using tools by itself, handling tough problems, and managing complex tasks—are key for better coding agents. Next, we plan to improve the model’s reasoning and decision-making, support more tasks, and update quickly based on how people use it.\n\n## Citation\n\nIf you find our work helpful feel free to give us a cite.\n\n```\n@techreport{qwen_qwen3_coder_next_tech_report,\n  title        = {Qwen3-Coder-Next Technical Report},\n  author       = {{Qwen Team}},\n  url          = {https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf},\n  note         = {Accessed: 2026-02-03}\n}\n```\n",
  "authors": [
    "QwenTeam"
  ],
  "captured_at": "2026-10-18T23:27:08.867657739+00:00",
  "category": "ai",
  "content_html": "<h2><a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf\">Tech Report</a> <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/QwenLM/Qwen3-Coder\">GitHub</a> <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://huggingface.co/collections/Qwen/qwen3-coder-next\">Hugging Face</a> <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://modelscope.cn/collections/Qwen/Qwen3-Coder-Next\">ModelScope</a> <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://discord.gg/yPEP2vHTu4\">DISCORD</a></h2>\n<h2>Introduction</h2>\n<p>We introduce <strong>Qwen3-Coder-Next</strong>, an open-weight language model designed specifically for coding agents and local development. Built on top of <strong>Qwen3-Next-80B-A3B-Base</strong>, which adopts a novel architecture with hybrid attention and MoE, Qwen3-Coder-Next has been agentically trained at scale on large-scale executable task synthesis, environment interaction, and reinforcement learning, obtaining strong coding and agentic capabilities with significantly lower inference costs.</p>\n<h2>Scaling Agentic Training</h2>\n<p>Rather than relying solely on parameter scaling, Qwen3-Coder-Next focuses on <strong>scaling agentic training signals</strong>. We train the model using large collections of <em>verifiable coding tasks</em> paired with executable environments, enabling the model to learn directly from environment feedback. This includes:</p>\n<ul>\n<li>Continued pretraining on code- and agent-centric data</li>\n<li>Supervised fine-tuning on data containing high-quality agent trajectories</li>\n<li>Domain-specialized expert training (e.g., software engineering, QA, web/UX)</li>\n<li>Expert distillation into a single deployment-ready model</li>\n</ul>\n<p>This recipe emphasizes <strong>long-horizon reasoning, tool usage, and recovery from execution failures</strong>, which are essential for real-world coding agents.</p>\n<h2>Performance on Coding Agent Benchmarks</h2>\n<h3>Agent-Centric Benchmark Results</h3>\n<p>The figure below summarizes performance across several widely used <strong>coding agent benchmarks</strong>, including SWE-Bench (Verified, Multilingual, and Pro), TerminalBench 2.0, and Aider.</p>\n<p><picture class=\"article-picture article-picture-fallback\" style=\"--image-ratio:16 / 9\"><img src=\"https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/benchmarks.png\" alt=\"Qwen3 Coder Next Benchmarks\" loading=\"eager\" decoding=\"async\" fetchpriority=\"high\" referrerpolicy=\"no-referrer\" class=\"progressive-image\"></picture></p>\n<p>The figure demonstrates that:</p>\n<ul>\n<li>Qwen3-Coder-Next achieves <strong>over 70% on SWE-Bench Verified</strong> using the SWE-Agent scaffold.</li>\n<li>Performance remains competitive across <strong>multilingual settings</strong> and the more challenging <strong>SWE-Bench Pro</strong> benchmark.</li>\n<li>Despite its small active footprint, the model matches or exceeds several much larger open-source models across agent-centric evaluations.</li>\n</ul>\n<p>As shown in the figure below, Our model achieves strong results on SWE-Bench Pro by scaling the number of agent turns, providing evidence that the model excels at long-horizon reasoning in multi-turn agentic tasks.</p>\n<p><picture class=\"article-picture article-picture-fallback\" style=\"--image-ratio:16 / 9\"><img src=\"https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro_vs_turns.png\" alt=\"SWE-Bench Pro turns\" loading=\"lazy\" decoding=\"async\" fetchpriority=\"low\" referrerpolicy=\"no-referrer\" class=\"progressive-image\"></picture></p>\n<h3>Efficiency–Performance Tradeoff</h3>\n<p>This figure highlights how Qwen3-Coder-Next achieves an improved Pareto tradeoff between efficiency and performance.</p>\n<p><picture class=\"article-picture article-picture-fallback\" style=\"--image-ratio:16 / 9\"><img src=\"https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro.png\" alt=\"Qwen3 Coder Next Main Image\" loading=\"lazy\" decoding=\"async\" fetchpriority=\"low\" referrerpolicy=\"no-referrer\" class=\"progressive-image\"></picture></p>\n<p>This comparison makes the efficiency story clear:</p>\n<ul>\n<li><strong>Qwen3-Coder-Next (3B active)</strong> achieves SWE-Bench-Pro performance comparable to models with <strong>10×–20× more active parameters</strong>.</li>\n<li>Qwen3-Coder-Next sits on a strong Pareto frontier for <strong>cost-effective agent deployment</strong>.</li>\n</ul>\n<h2>Demo</h2>\n<p>The small and fast coder model can be integrated into different downstream applications, and below we demonstrate showcases of OpenClaw, Qwen Code, Claude Code, Web dev, browser use, Cline, etc.</p>\n<h3>Web Dev</h3>\n<p>Creating a Chat Interface Next</p>\n<p>Create an Interactive Game Next</p>\n<p>Building an Interactive ASCII Art Drawing Tool Next</p>\n<h3>CLI</h3>\n<p>Desktop Cleanup Next</p>\n<p>Implementing a Web Game: Zombies vs. Plants Next</p>\n<h3>Cline</h3>\n<p>Creating a Multicolor Animation Next</p>\n<p>Building an Interactive Sound Art Tool Next</p>\n<h3>OpenClaw</h3>\n<p>Web Page for Qwen3 Coder Next Next</p>\n<h3>Browser Use Agent</h3>\n<p>Searching for a Product on Amazon Next</p>\n<p>Testing a Website Next</p>\n<h3>coder.qwen.ai</h3>\n<h2>Summary and Future Work</h2>\n<p>Qwen3-Coder-Next shows promising results on coding agent benchmarks, providing good speed and reasoning abilities for practical use. While it performs competitively—even compared to some larger open-source models—there is still much room for improvement.</p>\n<p>Looking ahead, we believe strong agent skills—like using tools by itself, handling tough problems, and managing complex tasks—are key for better coding agents. Next, we plan to improve the model’s reasoning and decision-making, support more tasks, and update quickly based on how people use it.</p>\n<h2>Citation</h2>\n<p>If you find our work helpful feel free to give us a cite.</p>\n<pre><code><span class=\"code-snippet\">@techreport{qwen_qwen3_coder_next_tech_report,\n  title        = {Qwen3-Coder-Next Technical Report},\n  author       = {{Qwen Team}},\n  url          = {https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf},\n  note         = {Accessed: 2026-02-03}\n}\n</span></code></pre>\n",
  "content_markdown": "## [Tech Report](https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf) [GitHub](https://github.com/QwenLM/Qwen3-Coder) [Hugging Face](https://huggingface.co/collections/Qwen/qwen3-coder-next) [ModelScope](https://modelscope.cn/collections/Qwen/Qwen3-Coder-Next) [DISCORD](https://discord.gg/yPEP2vHTu4)\n\n## Introduction\n\nWe introduce **Qwen3-Coder-Next**, an open-weight language model designed specifically for coding agents and local development. Built on top of **Qwen3-Next-80B-A3B-Base**, which adopts a novel architecture with hybrid attention and MoE, Qwen3-Coder-Next has been agentically trained at scale on large-scale executable task synthesis, environment interaction, and reinforcement learning, obtaining strong coding and agentic capabilities with significantly lower inference costs.\n\n## Scaling Agentic Training\n\nRather than relying solely on parameter scaling, Qwen3-Coder-Next focuses on **scaling agentic training signals**. We train the model using large collections of *verifiable coding tasks* paired with executable environments, enabling the model to learn directly from environment feedback. This includes:\n\n- Continued pretraining on code- and agent-centric data\n- Supervised fine-tuning on data containing high-quality agent trajectories\n- Domain-specialized expert training (e.g., software engineering, QA, web/UX)\n- Expert distillation into a single deployment-ready model\n\nThis recipe emphasizes **long-horizon reasoning, tool usage, and recovery from execution failures**, which are essential for real-world coding agents.\n\n## Performance on Coding Agent Benchmarks\n\n### Agent-Centric Benchmark Results\n\nThe figure below summarizes performance across several widely used **coding agent benchmarks**, including SWE-Bench (Verified, Multilingual, and Pro), TerminalBench 2.0, and Aider.\n\n![Qwen3 Coder Next Benchmarks](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/benchmarks.png)\n\nThe figure demonstrates that:\n\n- Qwen3-Coder-Next achieves **over 70% on SWE-Bench Verified** using the SWE-Agent scaffold.\n- Performance remains competitive across **multilingual settings** and the more challenging **SWE-Bench Pro** benchmark.\n- Despite its small active footprint, the model matches or exceeds several much larger open-source models across agent-centric evaluations.\n\nAs shown in the figure below, Our model achieves strong results on SWE-Bench Pro by scaling the number of agent turns, providing evidence that the model excels at long-horizon reasoning in multi-turn agentic tasks.\n\n![SWE-Bench Pro turns](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro_vs_turns.png)\n\n### Efficiency–Performance Tradeoff\n\nThis figure highlights how Qwen3-Coder-Next achieves an improved Pareto tradeoff between efficiency and performance.\n\n![Qwen3 Coder Next Main Image](https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen3-Coder-Next/swebench_pro.png)\n\nThis comparison makes the efficiency story clear:\n\n- **Qwen3-Coder-Next (3B active)** achieves SWE-Bench-Pro performance comparable to models with **10×–20× more active parameters**.\n- Qwen3-Coder-Next sits on a strong Pareto frontier for **cost-effective agent deployment**.\n\n## Demo\n\nThe small and fast coder model can be integrated into different downstream applications, and below we demonstrate showcases of OpenClaw, Qwen Code, Claude Code, Web dev, browser use, Cline, etc.\n\n### Web Dev\n\nCreating a Chat Interface Next\n\nCreate an Interactive Game Next\n\nBuilding an Interactive ASCII Art Drawing Tool Next\n\n### CLI\n\nDesktop Cleanup Next\n\nImplementing a Web Game: Zombies vs. Plants Next\n\n### Cline\n\nCreating a Multicolor Animation Next\n\nBuilding an Interactive Sound Art Tool Next\n\n### OpenClaw\n\nWeb Page for Qwen3 Coder Next Next\n\n### Browser Use Agent\n\nSearching for a Product on Amazon Next\n\nTesting a Website Next\n\n### coder.qwen.ai\n\n## Summary and Future Work\n\nQwen3-Coder-Next shows promising results on coding agent benchmarks, providing good speed and reasoning abilities for practical use. While it performs competitively—even compared to some larger open-source models—there is still much room for improvement.\n\nLooking ahead, we believe strong agent skills—like using tools by itself, handling tough problems, and managing complex tasks—are key for better coding agents. Next, we plan to improve the model’s reasoning and decision-making, support more tasks, and update quickly based on how people use it.\n\n## Citation\n\nIf you find our work helpful feel free to give us a cite.\n\n```\n@techreport{qwen_qwen3_coder_next_tech_report,\n  title        = {Qwen3-Coder-Next Technical Report},\n  author       = {{Qwen Team}},\n  url          = {https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf},\n  note         = {Accessed: 2026-02-03}\n}\n```\n",
  "dateCreated": "2026-10-18T23:27:08.867657739+00:00",
  "date_created": "2026-02-02T20:00:00+00:00",
  "date_modified": null,
  "date_published": "2026-02-02T20:00:00+00:00",
  "description": "Tech Report GitHub Hugging Face ModelScope DISCORD Introduction We introduce Qwen3-Coder-Next, an open-weight language model designed specifically for coding agents and local development. Built on top of Qwen3-Next-80B-A3B-Base, which…",
  "external_url": "https://qwen.ai/blog?id=qwen3-coder-next",
  "git": {
    "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e",
    "path": "items/qwen-ai/2026/02/2026-02-02-qwen3-coder-next-pushing-small-hybrid-models-on-agentic.md",
    "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/qwen-ai/2026/02/2026-02-02-qwen3-coder-next-pushing-small-hybrid-models-on-agentic.md"
  },
  "id": "urn:aggr:item:039520cd772f6367215a42793681c0e01e561de4",
  "isBasedOn": {
    "@id": "https://qwen.ai/blog?id=qwen3-coder-next"
  },
  "mainEntity": {
    "@id": "https://qwen.ai/blog?id=qwen3-coder-next",
    "@type": "CreativeWork",
    "archivedAt": {
      "@id": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-02-02-qwen3-coder-next-pushing-small-hybrid-models-on-agentic/#webpage"
    },
    "author": [
      {
        "@type": "Person",
        "name": "QwenTeam"
      }
    ],
    "datePublished": "2026-02-02T20:00:00+00:00",
    "headline": "Qwen3-Coder-Next: Pushing Small Hybrid Models on Agentic Coding",
    "image": {
      "@type": "ImageObject",
      "caption": null,
      "height": 90,
      "url": "https://img.alicdn.com/imgextra/i1/O1CN01LwK7vH21HjjuHvHWQ_!!6000000006960-2-tps-1590-954.png",
      "width": 160
    },
    "timeRequired": "PT3M",
    "url": "https://qwen.ai/blog?id=qwen3-coder-next",
    "wordCount": 587
  },
  "name": "Archived snapshot: Qwen3-Coder-Next: Pushing Small Hybrid Models on Agentic Coding",
  "preview": {
    "alt": null,
    "color": null,
    "height": 90,
    "placeholder": {
      "data_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=",
      "hash": ""
    },
    "url": "https://img.alicdn.com/imgextra/i1/O1CN01LwK7vH21HjjuHvHWQ_!!6000000006960-2-tps-1590-954.png",
    "width": 160
  },
  "reading_minutes": 3,
  "replicated_at": null,
  "source": "qwen-ai",
  "source_name": "Qwen",
  "summary": null,
  "tags": [
    "open-source"
  ],
  "temporalCoverage": "2026-10-18T23:27:08.867657739+00:00",
  "title": "Qwen3-Coder-Next: Pushing Small Hybrid Models on Agentic Coding",
  "url": "https://aggr.aymericbeaumet.com/items/qwen-ai/2026-02-02-qwen3-coder-next-pushing-small-hybrid-models-on-agentic/",
  "word_count": 587
}