开源项目 Hugging Face Models 2026-09-25T12:02 正文 ok 313 words

MiMo-V2.6-Distill-Qwen-9B:小米 MiMo 发布 90 亿参数 Agent 模型

英文标题:🤗 XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B

小米 MiMo 发布 MiMo-V2.6-Distill-Qwen-9B,这是一款基于 Qwen3.5-9B、使用 MiMo 生成数据进行监督微调(SFT)的 90 亿参数 Agent 模型。该模型覆盖代码、通用 Agent、视觉编程和网络安全任务,并作为 Agent 强化学习开放研究的初始检查点发布。官方评测显示,该模型在 SWE Pro、MiMo Code、MiMo Cyber、AutomationBench 等多项基准上显著超过 Qwen3.5-9B。开发者可通过支持 Qwen3.5 的 SGLang 部署模型,并使用 OpenAI 兼容 API 显式启用思维过程输出。
打开原文相关 1.6 · 机会 2
media4.5 KB media2.0 KB media411.9 KB

中文

MiMo-V2.6-Distill-Qwen-9B

MiMo-V2.6-Distill-Qwen-9B 是小米 MiMo 基于 MiMo 生成的数据,对 Qwen3.5-9B 进行监督微调开发的一款 90 亿参数 Agent 模型。它覆盖代码、通用 Agent 任务、视觉编程和网络安全领域。我们发布这一 SFT 检查点,作为 Agent 强化学习开放研究的起点。

评测

以下为 MiMo-V2.6 技术报告中披露的已发布 SFT 检查点评测结果。

领域 基准 指标 Qwen3.5-9B MiMo-V2.6-Distill-Qwen-9B(SFT)
代码 SWE Verified avg@3 60.0 61.1
代码 SWE Pro avg@3 32.0 44.6
代码 MiMo Code(mini)† avg@3 19.5 51.6
网络安全 MiMo Cyber(mini)† avg@3 5.7 31.3
通用 AutomationBench v1.0.6 avg@1 5.0 30.3
通用 Terminal Bench 2.1 avg@1 27.0 37.1
通用 Toolathlon-Verified avg@1 25.9 35.2
通用 OfficeQA avg@1 9.0 19.5
通用 JobBench avg@1 2.6 18.3
通用 MiMo General(mini)† avg@1 28.5 62.2
视觉 MiMo Visual Coding(mini)† avg@1 61.7 64.0

† 内部评测集。

训练数据

加权 SFT 数据混合物共包含 774 亿个 token,其中包括 272 亿个参与损失计算的 token。

领域 总 token 数(B) Token 占比(%) 参与损失计算的 token 数(B)
代码 23.2 29.9 7.3
网络安全 11.0 14.2 4.8
通用 22.0 28.5 5.7
视觉 21.2 27.4 9.4
总计 77.4 100.0 27.2

快速开始

如需进行文本生成,请使用近期发布且支持 Qwen3.5 的 SGLang 版本。该检查点包含其分词器以及 MiMo v2.6 聊天模板。

sglang serve \
  --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B \
  --reasoning-parser mimo \
  --host 0.0.0.0 \
  --port 30000

通过显式启用思维过程来请求该端点:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:30000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B",
    messages=[
        {"role": "user", "content": "What is 15% of 240?"}
    ],
    max_tokens=2048,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)

message = response.choices[0].message
print("Thinking:", getattr(message, "reasoning_content", "") or "")
print("Answer:", message.content or "")

引用

@misc{mimo2026v26,
  title={MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement},
  author={{Xiaomi MiMo Team}},
  year={2026},
  howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL}},
}

English / 原文

MiMo-V2.6-Distill-Qwen-9B

MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. It covers coding, general-purpose agent tasks, visual coding, and cybersecurity. We release this SFT checkpoint as a starting point for open research in agentic reinforcement learning.

Evaluation

Results for the released SFT checkpoint, as reported in the MiMo-V2.6 technical report.

Domain Benchmark Metric Qwen3.5-9B MiMo-V2.6-Distill-Qwen-9B (SFT)
Code SWE Verified avg@3 60.0 61.1
Code SWE Pro avg@3 32.0 44.6
Code MiMo Code (mini)† avg@3 19.5 51.6
Cyber MiMo Cyber (mini)† avg@3 5.7 31.3
General AutomationBench v1.0.6 avg@1 5.0 30.3
General Terminal Bench 2.1 avg@1 27.0 37.1
General Toolathlon-Verified avg@1 25.9 35.2
General OfficeQA avg@1 9.0 19.5
General JobBench avg@1 2.6 18.3
General MiMo General (mini)† avg@1 28.5 62.2
Visual MiMo Visual Coding (mini)† avg@1 61.7 64.0

† Internal evaluation sets.

Training Data

The weighted SFT data mixture contains 77.4B total tokens, including 27.2B loss-bearing tokens.

Domain Total tokens (B) Token share (%) Loss-bearing tokens (B)
Code 23.2 29.9 7.3
Cyber 11.0 14.2 4.8
General 22.0 28.5 5.7
Visual 21.2 27.4 9.4
Total 77.4 100.0 27.2

Quickstart

For text generation, use a recent SGLang build with Qwen3.5 support. The checkpoint includes its tokenizer and MiMo v2.6 chat template.

sglang serve \
  --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B \
  --reasoning-parser mimo \
  --host 0.0.0.0 \
  --port 30000

Query the endpoint with thinking explicitly enabled:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:30000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B",
    messages=[
        {"role": "user", "content": "What is 15% of 240?"}
    ],
    max_tokens=2048,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)

message = response.choices[0].message
print("Thinking:", getattr(message, "reasoning_content", "") or "")
print("Answer:", message.content or "")

Citation

@misc{mimo2026v26,
  title={MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement},
  author={{Xiaomi MiMo Team}},
  year={2026},
  howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL}},
}