{ "description": "Browse aymeric's, an independent, searchable archive of readable snapshots from followed feeds, preserved in Git with aggr.", "feed_url": "https://aggr.aymericbeaumet.com/sources/github.com/MoonshotAI/feed.json", "home_page_url": "https://aggr.aymericbeaumet.com/sources/github.com/MoonshotAI/", "items": [ { "content_html": "

We introduce Kimi-Dev-72B, our new open-source coding LLM for software engineering tasks. Kimi-Dev-72B achieves a new state-of-the-art on SWE-bench Verified among open-source models.

\n\n

\"Kimi

\n

Performance of Open-source Models on SWE-bench Verified.

\n

⚙️ Installation

\n
# clone repo\ngit clone https://github.com/MoonshotAI/Kimi-Dev.git\n# create env\nconda create -n kimidev python=3.12\n# local install\npip install -e .\n
\n

🛠️ How to use

\n

Prepare repo structure [From Agentless]

\n

Since for each issue in the benchmark (both SWE-Bench Lite and SWE-Bench Verified) we need to checkout the repository and process the files, you might want to save some time by downloading the preprocessed data here: swebench_repo_structure.zip. After downloading, please unzip and export the location as such

\n
export PROJECT_FILE_LOC={folder which you saved}\n
\n

Deploy vLLM Model

\n

Installation

\n
# Install vLLM with CUDA 12.8.\n# If you are using pip.\npip install vllm --extra-index-url https://download.pytorch.org/whl/cu128\n# If you are using uv.\nuv pip install vllm --torch-backend=auto\n
\n

Serving

\n
vllm serve Kimi-Dev-72B --served-model-name kimi-dev --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.95 --max-seq-len-to-capture 131072 --tensor-parallel-size 8\n
\n

Rollout

\n

Kimi-Dev adopts a simplified two-stage framework for handling code repair and test writing tasks:

\n
    \n
  1. File Localization: Intelligently identify key files that need modification based on problem descriptions and repository structure
    2. \n
  2. Code Editing: Perform precise code modifications on the located files, including bug fixes or unit test insertions
    3. \n
\n

Compared to multi-step localization methods, we perform localization at the file level and then pass the complete file to the repair step for more detailed reasoning.

\n

Run rollout script:

\n
conda activate kimidev\n# Bugfixer\npython kimidev/examples/rollout_messages_bugfixer.py --model_name {vllm_serve_model}\n# Testwriter\npython kimidev/examples/rollout_messages_testwriter.py --model_name {vllm_serve_model}\n
\n

👀 Example Results

\n

We provide some example result files as well as the files required for test-time scaling here.

\n

You can also download these files from Google Drive.

\n

💪 Contributing

\n

Welcome to submit Pull Requests or create Issues to help improve the project.

\n

😺 Contact

\n

If you have any questions, please feel free to submit a GitHub issue or contact zhuhan@moonshot.cn.

\n

📝 Citation

\n

If you find our code and models useful, please kindly cite the following information.

\n
@misc{yang2025kimidevagentlesstrainingskill,\n      title={Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents}, \n      author={Zonghan Yang and Shengjie Wang and Kelin Fu and Wenyang He and Weimin Xiong and Yibo Liu and Yibo Miao and Bofei Gao and Yejie Wang and Yingwei Ma and Yanhao Li and Yue Liu and Zhenxing Hu and Kaitai Zhang and Shuyi Wang and Huarong Chen and Flood Sung and Yang Liu and Yang Gao and Zhilin Yang and Tianyu Liu},\n      year={2025},\n      eprint={2509.23045},\n      archivePrefix={arXiv},\n      primaryClass={cs.AI},\n      url={https://arxiv.org/abs/2509.23045}, \n}\n
\n", "date_published": "2025-06-17T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Kimi-Dev", "id": "urn:aggr:item:5343bfc8c7dbbb043385827d672443c981572285", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71tea6dcmosb3tpn0e0?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-06-17", "tags": [ "ai" ], "title": "Kimi-Dev", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-06-17-kimi-dev/" }, { "content_html": "

\"\"

\n

Kimi-Audio-7B 🤗  | Kimi-Audio-7B-Instruct 🤗  | 📑 Paper

\n

We present Kimi-Audio, an open-source audio foundation model excelling in audio understanding, generation, and conversation. This repository contains the official implementation, models, and evaluation toolkit for Kimi-Audio.

\n

🔥🔥🔥 News!!

\n\n

Table of Contents

\n\n

Introduction

\n

Kimi-Audio is designed as a universal audio foundation model capable of handling a wide variety of audio processing tasks within a single unified framework. Key features include:

\n\n

Architecture Overview

\n

\"\"

\n

Kimi-Audio consists of three main components:

\n
    \n
  1. Audio Tokenizer: Converts input audio into:\n\n
    2. \n
  2. Audio LLM: A transformer-based model (initialized from a pre-trained text LLM like Qwen 2.5 7B) with shared layers processing multimodal inputs, followed by parallel heads for autoregressively generating text tokens and discrete audio semantic tokens.
    3. \n
  3. Audio Detokenizer: Converts the predicted discrete semantic audio tokens back into high-fidelity waveforms using a flow-matching model and a vocoder (BigVGAN), supporting chunk-wise streaming with a look-ahead mechanism for low latency.
    4. \n
\n

Getting Started

\n

Step1: Get the Code

\n
git clone https://github.com/MoonshotAI/Kimi-Audio.git\ncd Kimi-Audio\ngit submodule update --init --recursive\npip install -r requirements.txt\n
\n

Kimi‑Audio can now be installed directly via pip.

\n
pip install torch\npip install git+https://github.com/MoonshotAI/Kimi-Audio.git\n
\n

Quick Start

\n

This example demonstrates basic usage for generating text from audio (ASR) and generating both text and speech in a conversational turn.

\n
import soundfile as sf\nfrom kimia_infer.api.kimia import KimiAudio\n\n# --- 1. Load Model ---\nmodel_path = "moonshotai/Kimi-Audio-7B-Instruct" \nmodel = KimiAudio(model_path=model_path, load_detokenizer=True)\n\n# --- 2. Define Sampling Parameters ---\nsampling_params = {\n    "audio_temperature": 0.8,\n    "audio_top_k": 10,\n    "text_temperature": 0.0,\n    "text_top_k": 5,\n    "audio_repetition_penalty": 1.0,\n    "audio_repetition_window_size": 64,\n    "text_repetition_penalty": 1.0,\n    "text_repetition_window_size": 16,\n}\n\n# --- 3. Example 1: Audio-to-Text (ASR) ---\nmessages_asr = [\n    # You can provide context or instructions as text\n    {"role": "user", "message_type": "text", "content": "Please transcribe the following audio:"},\n    # Provide the audio file path\n    {"role": "user", "message_type": "audio", "content": "test_audios/asr_example.wav"}\n]\n\n# Generate only text output\n_, text_output = model.generate(messages_asr, **sampling_params, output_type="text")\nprint(">>> ASR Output Text: ", text_output) # Expected output: "这并不是告别，这是一个篇章的结束，也是新篇章的开始。"\n\n\n# --- 4. Example 2: Audio-to-Audio/Text Conversation ---\nmessages_conversation = [\n    # Start conversation with an audio query\n    {"role": "user", "message_type": "audio", "content": "test_audios/qa_example.wav"}\n]\n\n# Generate both audio and text output\nwav_output, text_output = model.generate(messages_conversation, **sampling_params, output_type="both")\n\n# Save the generated audio\noutput_audio_path = "output_audio.wav"\nsf.write(output_audio_path, wav_output.detach().cpu().view(-1).numpy(), 24000) # Assuming 24kHz output\nprint(f">>> Conversational Output Audio saved to: {output_audio_path}")\nprint(">>> Conversational Output Text: ", text_output) # Expected output: "当然可以，这很简单。一二三四五六七八九十。"\n\n# --- 5. Example 3: Audio-to-Audio/Text Conversation with Multiturn ---\n\nmessages = [\n    {"role": "user", "message_type": "audio", "content": "test_audios/multiturn/case2/multiturn_q1.wav"},\n    # This is the first turn output of Kimi-Audio\n    {"role": "assistant", "message_type": "audio-text", "content": ["test_audios/multiturn/case2/multiturn_a1.wav", "当然可以，这很简单。一二三四五六七八九十。"]},\n    {"role": "user", "message_type": "audio", "content": "test_audios/multiturn/case2/multiturn_q2.wav"}\n]\nwav, text = model.generate(messages, **sampling_params, output_type="both")\n\n\n# Generate both audio and text output\nwav_output, text_output = model.generate(messages_conversation, **sampling_params, output_type="both")\n\n# Save the generated audio\noutput_audio_path = "output_audio.wav"\nsf.write(output_audio_path, wav_output.detach().cpu().view(-1).numpy(), 24000) # Assuming 24kHz output\nprint(f">>> Conversational Output Audio saved to: {output_audio_path}")\nprint(">>> Conversational Output Text: ", text_output) # Expected output: "没问题，继续数下去就是十一十二十三十四十五十六十七十八十九二十。"\n\nprint("Kimi-Audio inference examples complete.")\n
\n

Evaluation

\n

Kimi-Audio achieves state-of-the-art (SOTA) performance across a wide range of audio benchmarks.

\n

The below is the overall performance:

\n

\"\"

\n

Here are performances on different benchmarks, you can easily reproduce the our results and baselines by our Kimi-Audio-Evalkit (also see Evaluation Toolkit):

\n

Automatic Speech Recognition (ASR)

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
DatasetsModelPerformance (WER↓)
LibriSpeech test-clean | test-otherQwen2-Audio-base1.74 | 4.04
Baichuan-base3.02 | 6.04
Step-Audio-chat3.19 | 10.67
Qwen2.5-Omni2.37 | 4.21
Kimi-Audio1.28 | 2.42
Fleurs zh | enQwen2-Audio-base3.63 | 5.20
Baichuan-base4.15 | 8.07
Step-Audio-chat4.26 | 8.56
Qwen2.5-Omni2.92 | 4.17
Kimi-Audio2.69 | 4.44
AISHELL-1Qwen2-Audio-base1.52
Baichuan-base1.93
Step-Audio-chat2.14
Qwen2.5-Omni1.13
Kimi-Audio0.60
AISHELL-2 iosQwen2-Audio-base3.08
Baichuan-base3.87
Step-Audio-chat3.89
Qwen2.5-Omni2.56
Kimi-Audio2.56
WenetSpeech test-meeting | test-netQwen2-Audio-base8.40 | 7.64
Baichuan-base13.28 | 10.13
Step-Audio-chat10.83 | 9.47
Qwen2.5-Omni7.71 | 6.04
Kimi-Audio6.28 | 5.37
Kimi-ASR Internal Testset subset1 | subset2Qwen2-Audio-base2.31 | 3.24
Baichuan-base3.41 | 5.60
Step-Audio-chat2.82 | 4.74
Qwen2.5-Omni1.53 | 2.68
Kimi-Audio1.42 | 2.44
\n

Audio Understanding

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
DatasetsModelPerformance↑
MMAU music | sound | speechQwen2-Audio-base58.98 | 69.07 | 52.55
Baichuan-chat49.10 | 59.46 | 42.47
GLM-4-Voice38.92 | 43.54 | 32.43
Step-Audio-chat49.40 | 53.75 | 47.75
Qwen2.5-Omni62.16 | 67.57 | 53.92
Kimi-Audio61.68 | 73.27 | 60.66
ClothoAQA test | devQwen2-Audio-base71.73 | 72.63
Baichuan-chat48.02 | 48.16
Step-Audio-chat45.84 | 44.98
Qwen2.5-Omni72.86 | 73.12
Kimi-Audio71.24 | 73.18
VocalSoundQwen2-Audio-base93.82
Baichuan-base58.17
Step-Audio-chat28.58
Qwen2.5-Omni93.73
Kimi-Audio94.85
Nonspeech7kQwen2-Audio-base87.17
Baichuan-chat59.03
Step-Audio-chat21.38
Qwen2.5-Omni69.89
Kimi-Audio93.93
MELDQwen2-Audio-base51.23
Baichuan-chat23.59
Step-Audio-chat33.54
Qwen2.5-Omni49.83
Kimi-Audio59.13
TUT2017Qwen2-Audio-base33.83
Baichuan-base27.9
Step-Audio-chat7.41
Qwen2.5-Omni43.27
Kimi-Audio65.25
CochlScene test | devQwen2-Audio-base52.69 | 50.96
Baichuan-base34.93 | 34.56
Step-Audio-chat10.06 | 10.42
Qwen2.5-Omni63.82 | 63.82
Kimi-Audio79.84 | 80.99
\n

Audio-to-Text Chat

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
DatasetsModelPerformance↑
OpenAudioBench AlpacaEval | Llama Questions | Reasoning QA | TriviaQA | Web QuestionsQwen2-Audio-chat57.19 | 69.67 | 42.77 | 40.30 | 45.20
Baichuan-chat59.65 | 74.33 | 46.73 | 55.40 | 58.70
GLM-4-Voice57.89 | 76.00 | 47.43 | 51.80 | 55.40
StepAudio-chat56.53 | 72.33 | 60.00 | 56.80 | 73.00
Qwen2.5-Omni72.76 | 75.33 | 63.76 | 57.06 | 62.80
Kimi-Audio75.73 | 79.33 | 58.02 | 62.10 | 70.20
VoiceBench AlpacaEval | CommonEval | SD-QA | MMSUQwen2-Audio-chat3.69 | 3.40 | 35.35 | 35.43
Baichuan-chat4.00 | 3.39 | 49.64 | 48.80
GLM-4-Voice4.06 | 3.48 | 43.31 | 40.11
StepAudio-chat3.99 | 2.99 | 46.84 | 28.72
Qwen2.5-Omni4.33 | 3.84 | 57.41 | 56.38
Kimi-Audio4.46 | 3.97 | 63.12 | 62.17
VoiceBench OpenBookQA | IFEval | AdvBench | AvgQwen2-Audio-chat49.01 | 22.57 | 98.85 | 54.72
Baichuan-chat63.30 | 41.32 | 86.73 | 62.51
GLM-4-Voice52.97 | 24.91 | 88.08 | 57.17
StepAudio-chat31.87 | 29.19 | 65.77 | 48.86
Qwen2.5-Omni79.12 | 53.88 | 99.62 | 72.83
Kimi-Audio83.52 | 61.10 | 100.00 | 76.93
\n

Speech Conversation

\n

Performance of Kimi-Audio and baseline models on speech conversation.

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelAbility
GPT-4o4.213.654.053.874.544.06
Step-Audio-chat3.252.873.333.054.143.33
GLM-4-Voice3.833.513.773.074.043.65
GPT-4o-mini3.152.714.243.164.013.45
Kimi-Audio4.303.454.273.394.093.90
\n

Finetune

\n

We release the pre-trained model and the lightweight finetune codes. Please refer to the finetune_codes/README.md for more details.

\n

Evaluation Toolkit

\n

Evaluating and comparing audio foundation models is challenging due to inconsistent metrics, varying inference configurations, and a lack of standardized generation evaluation. To address this, we developed and open-sourced an Evaluation Toolkit.

\n

Key features:

\n\n

We encourage the community to use and contribute to this toolkit to foster more reliable and comparable benchmarking. Find it here: Kimi-Audio-Evalkit.

\n

Generation Testset

\n

We collect and release Kimi-Audio-Generation-Testset, which is designed to benchmark and evaluate the conversational capabilities of audio-based dialogue models. It consists of a collection of audio files containing various instructions and conversational prompts. The primary goal is to assess a model's ability to generate not just relevant, but also appropriately styled audio responses. The language in dataset is Chinese.

\n

License

\n

The model is based and modified from Qwen 2.5-7B. Code derived from Qwen2.5-7B is licensed under the Apache 2.0 License. Other parts of the code are licensed under the MIT License.

\n

Acknowledgements

\n

We would like to thank the following projects and individuals for their contributions to the development of Kimi-Audio:

\n\n

Thank you to all the open-source projects for their contributions to this project!

\n

Citation

\n

If you find Kimi-Audio useful in your research or applications, please cite our technical report:

\n
@misc{kimiteam2025kimiaudiotechnicalreport,\n      title={Kimi-Audio Technical Report}, \n      author={KimiTeam and Ding Ding and Zeqian Ju and Yichong Leng and Songxiang Liu and Tong Liu and Zeyu Shang and Kai Shen and Wei Song and Xu Tan and Heyi Tang and Zhengtao Wang and Chu Wei and Yifei Xin and Xinran Xu and Jianwei Yu and Yutao Zhang and Xinyu Zhou and Y. Charles and Jun Chen and Yanru Chen and Yulun Du and Weiran He and Zhenxing Hu and Guokun Lai and Qingcheng Li and Yangyang Liu and Weidong Sun and Jianzhou Wang and Yuzhi Wang and Yuefeng Wu and Yuxin Wu and Dongchao Yang and Hao Yang and Ying Yang and Zhilin Yang and Aoxiong Yin and Ruibin Yuan and Yutong Zhang and Zaida Zhou},\n      year={2025},\n      eprint={2504.18425},\n      archivePrefix={arXiv},\n      primaryClass={eess.AS},\n      url={https://arxiv.org/abs/2504.18425}, \n}\n
\n

Contact Us

\n

For questions, issues, or collaboration inquiries, please feel free to open an issue on GitHub.

\n", "date_published": "2025-04-26T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Kimi-Audio", "id": "urn:aggr:item:ec8df625def3c9440551795eb8a41a9109ac6c46", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71teed3v89kkegmsvo0?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-04-26", "tags": [ "ai" ], "title": "Kimi-Audio", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-04-26-kimi-audio/" }, { "content_html": "

📈 Introducing Kimina-Prover Preview, the first large formal reasoning model that can reason in a human-like way and prove mathematical theorems rigorously in the Lean 4 language.

\n

Some key ingredients about the design and training of Kimina-Prover Preview are listed as follows.

\n
@article{kimina_prover_2025,\n\ttitle = {Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning},\n\tauthor = {Wang, Haiming and Unsal, Mert and Lin, Xiaohan and Baksys, Mantas and Liu, Junqi and Santos, Marco Dos and Sung, Flood and Vinyes, Marina and Ying, Zhenzhe and Zhu, Zekai and Lu, Jianqiao and Saxcé, Hugues de and Bailey, Bolton and Song, Chendong and Xiao, Chenjun and Zhang, Dehao and Zhang, Ebony and Pu, Frederick and Zhu, Han and Liu, Jiawei and Bayer, Jonas and Michel, Julien and Yu, Longhui and Dreyfus-Schmidt, Léo and Tunstall, Lewis and Pagani, Luigi and Machado, Moreira and Bourigault, Pauline and Wang, Ran and Polu, Stanislas and Barroyer, Thibaut and Li, Wen-Ding and Niu, Yazhe and Fleureau, Yann and Hu, Yangyang and Yu, Zhouliang and Wang, Zihan and Yang, Zhilin and Liu, Zhengying and Li, Jia},\n\tyear = {2025},\n\turl = {http://arxiv.org/abs/2504.11354},\n}\n
\n", "date_published": "2025-04-15T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Kimina-Prover-Preview", "id": "urn:aggr:item:32b3ea95359e7a55dfc7dd43d0cfc83a0cc0f0e9", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71tein6rtp4tq9po6h0?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-04-15", "tags": [ "ai" ], "title": "Kimina-Prover Preview", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-04-15-kimina-prover-preview/" }, { "content_html": "

1. Introduction

\n

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities—all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B).

\n

Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent interaction tasks (e.g.,OSWorld), achieving state-of-the-art results comparable to flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, optical character recognition (OCR), mathematical reasoning, multi-image understanding, and etc.

\n

In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several specialized domains.

\n

Kimi-VL also advances the pareto frontiers of multimodal models in processing long contexts and perceiving clearly: Equipped with a 128K extended context window, Kimi-VL can processes long and diverse inputs, achieving impressive scores of 64.5 on LongVideoBench, and 35.1 on MMLongBench-Doc; Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost with common visual inputs and general tasks.

\n

Building on this foundation, we introduce an advanced long-thinking variant: Kimi-VL-Thinking. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), this model exhibits strong long-horizon reasoning capabilities. It achieves scores of 61.7 on MMMU, 36.8 on MathVision, and 71.3 on MathVista while maintaining the compact 2.8B activated LLM parameter footprint, setting a new standard for efficient yet capable multimodal thinking models.

\n

Besides original model variants, we also provide a new Kimi-VL-A3B-Thinking-2506 variant with several new or improved abilities:

\n\n

2. Architecture

\n

The model adopts an MoE language model, a native-resolution visual encoder (MoonViT), and an MLP projector, as illustrated in the following image.

\n

\"\"

\n

3. News

\n\n

4. Model Variants

\n

🤗 For common general multimodal perception and understanding, OCR, long video and long document, video perception, and OS-agent uses, we recommend Kimi-VL-A3B-Instruct for efficient inference; meanwhile, our new thinking version, Kimi-VL-A3B-Thinking-2506 also has excellent multimodal perception, long video and long document and OS-agent grounding abilities while achieving better multimodal reasoning skills. See this blog for more information.

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Model#Total Params#Activated ParamsContext LengthDownload Link
🔥Kimi-VL-A3B-Thinking-250616B3B128K🤗 Hugging Face
Kimi-VL-A3B-Instruct16B3B128K🤗 Hugging Face
Kimi-VL-A3B-Thinking (deprecated)16B3B128K🤗 Hugging Face
\n

Note

\n

Recommended parameter settings:

\n\n

Hugging Face Demo

\n
\n

🤗 We serve our model demo in Hugging Face spaces:

\n\n
\n

5. Performance

\n

Note

\n

See the performance of Kimi-VL-A3B-Thinking-2506 at Hugging Face.

\n

As an efficient model, Kimi-VL can robustly handle diverse tasks (fine-grained perception, math, college-level problems, OCR, agent, etc) across a broad spectrum of input forms (single-image, multi-image, video, long-document, etc).

\n

A brief comparison with existing 10B-level dense VLMs and DeepSeek-VL2 (A4.5B):

\n

\"\"

\n

With effective long-thinking abilities, Kimi-VL-A3B-Thinking (2504 version) can match the performance of 30B/70B frontier open-source VLMs on MathVision benchmark:

\n

\"\"

\n

6. Example usage

\n

Setup

\n
conda create -n kimi-vl python=3.10 -y\nconda activate kimi-vl\npip install -r requirements.txt\n
\n

Note

\n

If you encounter Out-of-Memory or want to speed up inference, please install flash-attn with pip install flash-attn --no-build-isolation.

\n

Inference with Hugging Face Transformers

\n

We introduce how to use our model at inference stage using transformers library. It is recommended to use python=3.10, torch=2.5.1, and transformers=4.51.3 as the development environment.

\n

Kimi-VL-A3B-Instruct:

\n
import torch\nfrom PIL import Image\nfrom transformers import AutoModelForCausalLM, AutoProcessor\n\nmodel_path = "moonshotai/Kimi-VL-A3B-Instruct"\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_path,\n    torch_dtype="auto",\n    device_map="auto",\n    trust_remote_code=True,\n)\n# If flash-attn has been installed, it is recommended to set torch_dtype=torch.bfloat16 and attn_implementation="flash_attention_2"\n# to save memory and speed up inference\n# model = AutoModelForCausalLM.from_pretrained(\n#     model_path,\n#     torch_dtype=torch.bfloat16,\n#     device_map="auto",\n#     trust_remote_code=True,\n#     attn_implementation="flash_attention_2"\n# )\n\nprocessor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)\n\nimage_path = "./figures/demo.png"\nimage = Image.open(image_path)\nmessages = [\n    {"role": "user", "content": [{"type": "image", "image": image_path}, {"type": "text", "text": "What is the dome building in the picture? Think step by step."}]}\n]\ntext = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")\ninputs = processor(images=image, text=text, return_tensors="pt", padding=True, truncation=True).to(model.device)\ngenerated_ids = model.generate(**inputs, max_new_tokens=512)\ngenerated_ids_trimmed = [\n    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)\n]\nresponse = processor.batch_decode(\n    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False\n)[0]\nprint(response)\n
\n

Kimi-VL-A3B-Thinking-2506:

\n
import torch\nfrom PIL import Image\nfrom transformers import AutoModelForCausalLM, AutoProcessor\n\nmodel_path = "moonshotai/Kimi-VL-A3B-Thinking-2506"\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_path,\n    torch_dtype="auto",\n    device_map="auto",\n    trust_remote_code=True,\n)\n# If flash-attn has been installed, it is recommended to set torch_dtype=torch.bfloat16 and attn_implementation="flash_attention_2"\n# to save memory and speed up inference\n# model = AutoModelForCausalLM.from_pretrained(\n#     model_path,\n#     torch_dtype=torch.bfloat16,\n#     device_map="auto",\n#     trust_remote_code=True,\n#     attn_implementation="flash_attention_2"\n# )\nprocessor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)\n\nimage_paths = ["./figures/demo1.png", "./figures/demo2.png"]\nimages = [Image.open(path) for path in image_paths]\nmessages = [\n    {\n        "role": "user",\n        "content": [\n            {"type": "image", "image": image_path} for image_path in image_paths\n        ] + [{"type": "text", "text": "Please infer step by step who this manuscript belongs to and what it records"}],\n    },\n]\ntext = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")\ninputs = processor(images=images, text=text, return_tensors="pt", padding=True, truncation=True).to(model.device)\ngenerated_ids = model.generate(**inputs, max_new_tokens=32768, temperature=0.8)\ngenerated_ids_trimmed = [\n    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)\n]\nresponse = processor.batch_decode(\n    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False\n)[0]\nprint(response)\n
\n

7. Finetuning

\n

Collaborating closely with the open-source community, Kimi-VL now offers seamless support for efficient fine-tuning through the latest version of LLaMA-Factory.

\n

The framework enables Single-GPU LoRA fine-tuning with 50GB of VRAM, as well as Multi-GPU full/lora fine-tuning using DeepSpeed ZeRO-2. For more detailed configuration instructions, check out this PR.

\n

8. Deployment

\n

Using vLLM

\n

The vLLM main branch has supported Kimi-VL deployment. You are welcome to deploy Kimi-VL using vLLM.

\n

Offline Inference

\n
from PIL import Image\nfrom transformers import AutoProcessor\nfrom vllm import LLM, SamplingParams\n\nmodel_path = "moonshotai/Kimi-VL-A3B-Instruct"  # or "moonshotai/Kimi-VL-A3B-Thinking-2506"\nllm = LLM(\n    model_path,\n    trust_remote_code=True,\n)\n\nprocessor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)\n\nimage_path = "./figures/demo.png"\nimage = Image.open(image_path)\nmessages = [\n    {"role": "user", "content": [{"type": "image", "image": image_path}, {"type": "text", "text": "What is the dome building in the picture? Think step by step."}]}\n]\ntext = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")\noutputs = llm.generate([{"prompt": text, "multi_modal_data": {"image": image}}], sampling_params = SamplingParams(max_tokens=512))\n\nprint("-" * 50)\nfor o in outputs:\n    generated_text = o.outputs[0].text\n    print(generated_text)\n    print("-" * 50)\n
\n

OpenAI-Compatible Server

\n

Serve Kimi-VL with vllm serve command:

\n
# If you need a longer context window, you can set --max-model-len and --max-num-batched-tokens to 131072\n# If you need more input images, you can set --limit-mm-per-prompt image=256 or 512\n\n# kimi-vl-thinking-2506\nvllm serve moonshotai/Kimi-VL-A3B-Thinking-2506 --served-model-name kimi-vl-thinking-2506 --trust-remote-code --tensor-parallel-size 1 --max-num-batched-tokens 32768 --max-model-len 32768 --limit-mm-per-prompt image=64\n\n# kimi-vl-instruct\nvllm serve moonshotai/Kimi-VL-A3B-Instruct --served-model-name kimi-vl --trust-remote-code --tensor-parallel-size 1 --max-num-batched-tokens 32768 --max-model-len 32768 --limit-mm-per-prompt image=64\n
\n

Call the API

\n
import base64\nfrom PIL import Image\nfrom io import BytesIO\nfrom openai import OpenAI\n\nclient = OpenAI(\n    base_url="http://localhost:8000/v1",\n    api_key="token-abc123",\n)\n\nimage_path = "./figures/demo.png"\nimage = Image.open(image_path).convert("RGB")\n\nbuffered = BytesIO()\nimage.save(buffered, format="JPEG")\nimg_b64_str = base64.b64encode(buffered.getvalue()).decode("utf-8")\nbase64_image_url = f"data:image/jpeg;base64,{img_b64_str}"\n\nmessages = [\n    {"role": "user", "content": [{"type": "image_url", "image_url": {"url": base64_image_url}}, {"type": "text", "text": "What is the dome building in the picture? Think step by step."}]}\n]\n\ncompletion = client.chat.completions.create(\n  model="kimi-vl-thinking-2506", # or kimi-vl\n  messages=messages\n)\n\nprint(completion.choices[0].message)\n
\n

9. Citation

\n
@misc{kimiteam2025kimivltechnicalreport,\n      title={{Kimi-VL} Technical Report}, \n      author={Kimi Team and Angang Du and Bohong Yin and Bowei Xing and Bowen Qu and Bowen Wang and Cheng Chen and Chenlin Zhang and Chenzhuang Du and Chu Wei and Congcong Wang and Dehao Zhang and Dikang Du and Dongliang Wang and Enming Yuan and Enzhe Lu and Fang Li and Flood Sung and Guangda Wei and Guokun Lai and Han Zhu and Hao Ding and Hao Hu and Hao Yang and Hao Zhang and Haoning Wu and Haotian Yao and Haoyu Lu and Heng Wang and Hongcheng Gao and Huabin Zheng and Jiaming Li and Jianlin Su and Jianzhou Wang and Jiaqi Deng and Jiezhong Qiu and Jin Xie and Jinhong Wang and Jingyuan Liu and Junjie Yan and Kun Ouyang and Liang Chen and Lin Sui and Longhui Yu and Mengfan Dong and Mengnan Dong and Nuo Xu and Pengyu Cheng and Qizheng Gu and Runjie Zhou and Shaowei Liu and Sihan Cao and Tao Yu and Tianhui Song and Tongtong Bai and Wei Song and Weiran He and Weixiao Huang and Weixin Xu and Xiaokun Yuan and Xingcheng Yao and Xingzhe Wu and Xinxing Zu and Xinyu Zhou and Xinyuan Wang and Y. Charles and Yan Zhong and Yang Li and Yangyang Hu and Yanru Chen and Yejie Wang and Yibo Liu and Yibo Miao and Yidao Qin and Yimin Chen and Yiping Bao and Yiqin Wang and Yongsheng Kang and Yuanxin Liu and Yulun Du and Yuxin Wu and Yuzhi Wang and Yuzi Yan and Zaida Zhou and Zhaowei Li and Zhejun Jiang and Zheng Zhang and Zhilin Yang and Zhiqi Huang and Zihao Huang and Zijia Zhao and Ziwei Chen},\n      year={2025},\n      eprint={2504.07491},\n      archivePrefix={arXiv},\n      primaryClass={cs.CV},\n      url={https://arxiv.org/abs/2504.07491}, \n}\n
\n", "date_published": "2025-04-10T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Kimi-VL", "id": "urn:aggr:item:7fc289291c7ec22f59bafba1cbea421ad4b7442a", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71tdlmdcmosb3tpmsig?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-04-10", "tags": [ "ai" ], "title": "Kimi-VL", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-04-10-kimi-vl/" }, { "content_html": "

\"\"

\n

Abstract

\n

Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves ∼ 2× computational efficiency compared to AdamW with compute optimal training.

\n

Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models.

\n

We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research.

\n

Our code is available at MoonshotAI/Moonlight.

\n

Key Ingredients

\n

Our work builds upon Muon while systematically identifying and resolving its limitations in large-scale training scenarios. Our technical contributions include:

\n\n

\"\"

\n

Scaling up with Muon. (a) Scaling law experiments comparing Muon and Adam. Muon is 2 times more sample efficient than Adam. (b) The MMLU performance of our Moonlight model optimized with Muon and other comparable models. Moonlight advances the Pareto frontier of performance vs training FLOPs.

\n

Performance

\n

We named our lightweight model trained with Muon "Moonlight". We compared Moonlight with SOTA public models at similar scale:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Benchmark (Metric)Llama3.2-3BQwen2.5-3BDSV2-LiteMoonlight
Activated Param†2.81B2.77B2.24B2.24B
Total Params†2.81B2.77B15.29B15.29B
Training Tokens9T18T5.7T5.7T
OptimizerAdamW*AdamWMuon
EnglishMMLU54.7565.658.370.0
MMLU-pro25.034.625.542.4
BBH46.856.344.165.2
TriviaQA‡59.651.165.166.3
CodeHumanEval28.042.129.948.1
MBPP48.757.143.263.8
MathGSM8K34.079.141.177.4
MATH8.542.617.145.3
CMath-80.058.481.1
ChineseC-Eval-75.060.377.2
CMMLU-75.064.378.2
\n

Qwen 2 & 2.5 reports didn't disclose their optimizer information. †The reported parameter counts exclude the embedding parameters. ‡We test all listed models with the full set of TriviaQA.

\n

Example usage

\n

Model Download

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Model#Total Params#Activated ParamsContext LengthDownload Link
Moonlight16B3B8K🤗 Hugging Face
Moonlight-Instruct16B3B8K🤗 Hugging Face
\n

Inference with Hugging Face Transformers

\n

We introduce how to use our model at inference stage using transformers library. It is recommended to use python=3.10, torch>=2.1.0, and transformers=4.48.2 as the development environment.

\n

For our pretrained model (Moonlight):

\n
from transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_path = "moonshotai/Moonlight-16B-A3B"\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_path,\n    torch_dtype="auto",\n    device_map="auto",\n    trust_remote_code=True,\n)\ntokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)\n\nprompt = "1+1=2, 1+2="\ninputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True).to(model.device)\ngenerated_ids = model.generate(**inputs, max_new_tokens=100)\nresponse = tokenizer.batch_decode(generated_ids)[0]\nprint(response)\n
\n

For our instruct model (Moonlight-Instruct):

\n
from transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_path = "moonshotai/Moonlight-16B-A3B-Instruct"\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_path,\n    torch_dtype="auto",\n    device_map="auto",\n    trust_remote_code=True\n)\ntokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)\n\nmessages = [\n    {"role": "system", "content": "You are a helpful assistant provided by Moonshot-AI."},\n    {"role": "user", "content": "Is 123 a prime?"}\n]\ninput_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)\ngenerated_ids = model.generate(inputs=input_ids, max_new_tokens=500)\nresponse = tokenizer.batch_decode(generated_ids)[0]\nprint(response)\n
\n

Moonlight has the same architecture as DeepSeek-V3, which is supported by many popular inference engines, such as VLLM and SGLang. As a result, our model can also be easily deployed using these tools.

\n

Training

\n
# train qwen-like dense model with muon\npython3 examples/toy_train.py --model qwen --optimizer muon --dataset openwebtext-100k --hidden_size 896 --lr 1e-3\n\n# train qwen-like dense model with adamw\npython3 examples/toy_train.py --model qwen --optimizer adamw --dataset openwebtext-100k --hidden_size 896 --lr 1e-3\n
\n

Intermediate Checkpoints

\n

To support ongoing research efforts, we will soon release our intermediate checkpoints. Coming soon...

\n

Citation

\n

If you find Moonlight is useful or want to use in your projects, please kindly cite our paper:

\n
@misc{liu2025muonscalablellmtraining,\n      title={Muon is Scalable for LLM Training}, \n      author={Jingyuan Liu and Jianlin Su and Xingcheng Yao and Zhejun Jiang and Guokun Lai and Yulun Du and Yidao Qin and Weixin Xu and Enzhe Lu and Junjie Yan and Yanru Chen and Huabin Zheng and Yibo Liu and Shaowei Liu and Bohong Yin and Weiran He and Han Zhu and Yuzhi Wang and Jianzhou Wang and Mengnan Dong and Zheng Zhang and Yongsheng Kang and Hao Zhang and Xinran Xu and Yutao Zhang and Yuxin Wu and Xinyu Zhou and Zhilin Yang},\n      year={2025},\n      eprint={2502.16982},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2502.16982}, \n}\n
\n", "date_published": "2025-02-23T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Moonlight", "id": "urn:aggr:item:79b48f27241a8dfbf4acb69eb367cfc8f8b1eb26", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71teo1l51jas5dtnj8g?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-02-23", "tags": [ "ai" ], "title": "Muon is Scalable for LLM Training", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-02-23-muon-is-scalable-for-llm-training/" }, { "content_html": "

\"\"

\n

\"\" Full Report

\n

🚀 Introducing MoBA --- Mixture of Block Attention

\n\n

\"\" \"\"

\n
\n

Note: MoBA requires continue training of existing models to achieve its acceleration benefits. It is not a drop-in sparse attention solution that can be directly applied to pretrained models without additional training.

\n
\n

Abstract

\n

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored.

\n

In this work, we propose a solution that adheres to the “less structure” principle, allowing the model to autonomously determine where to attend, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi’s long-context requests and demonstrates significant advancements in efficient attention computation for LLMs.

\n

Our code is available at MoonshotAI/MoBA.

\n

\"\"

\n

Evaluation with 1M context length

\n

\"\"

\n

Environment Setup

\n

Note that current kernel implementations rely on flash-attn==2.6.3 and torch >= 2.1.0

\n
conda create -n moba python=3.10\nconda activate moba\npip install .\n
\n

Quick Start

\n

We provide a transformers-friendly implementation for MoBA.

\n

Feel free to choose attention backends by --attn between moba and moba_naive.

\n
python3 examples/llama.py --model meta-llama/Llama-3.1-8B --attn moba\n
\n

Implementation Details

\n\n

Unit Tests

\n
pytest tests/test_moba_attn.py\n
\n

References

\n\n

Citation

\n

If you find MoBA is useful or want to use in your projects, please kindly cite our paper:

\n
@article{lu2025mobamixtureblockattention,\n  author = {Enzhe Lu and Zhejun Jiang and Jingyuan Liu and Yulun Du and Tao Jiang and Chao Hong and Shaowei Liu and Weiran He and Enming Yuan and Yuzhi Wang and Zhiqi Huang and Huan Yuan and Suting Xu and Xinran Xu and Guokun Lai and Yanru Chen and Huabin Zheng and Junjie Yan and Jianlin Su and Yuxin Wu and Yutao Zhang and Zhilin Yang and Xinyu Zhou and Mingxing Zhang and Jiezhong Qiu},\n  title = {MoBA: Mixture of Block Attention for Long-Context LLMs},\n  journal={arXiv preprint arXiv:2502.13189},\n  year={2025}\n}\n
\n", "date_published": "2025-02-18T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/MoBA", "id": "urn:aggr:item:8b6a8b0f3d3f7dba3bdc5702272342718b46651d", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71tet1l51jas5dtnjr0?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-02-18", "tags": [ "ai" ], "title": "MoBA: Mixture of Block Attention for Long-Context LLMs", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-02-18-moba-mixture-of-block-attention-for-long-context-llms/" }, { "content_html": "

There are a few key ingredients about the design and training of k1.5.

\n
@article{team2025kimi,\n  title={Kimi k1.5: Scaling reinforcement learning with llms},\n  author={Team, Kimi and Du, Angang and Gao, Bofei and Xing, Bowei and Jiang, Changjiu and Chen, Cheng and Li, Cheng and Xiao, Chenjun and Du, Chenzhuang and Liao, Chonghua and others},\n  journal={arXiv preprint arXiv:2501.12599},\n  year={2025}\n}\n
\n", "date_published": "2025-01-20T00:00:00+00:00", "external_url": "https://github.com/MoonshotAI/Kimi-k1.5", "id": "urn:aggr:item:77214df8bd6f2f4b0fe92ab07043216b2a5a9ad9", "image": "https://kimi-file.kimi.ai/prod-chat-kimi/kfs/4/2/2026-03-25/1d71tf1qav1fc647s9li0?x-tos-process=image%2Fauto-orient%2C1%2Fstrip%2Fignore-error%2C1", "language": "en-CN", "summary": "2025-01-20", "tags": [ "ai" ], "title": "Kimi K1.5: Scaling Reinforcement Learning with LLMs", "url": "https://aggr.aymericbeaumet.com/items/kimi-com/2025-01-20-kimi-k1-5-scaling-reinforcement-learning-with-llms/" } ], "language": "en", "title": "github.com/MoonshotAI", "version": "https://jsonfeed.org/version/1.1" }