Xiaomi MiMo released MiMo-V2.6-Distill-Qwen-9B on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Distill-Qwen-9B, a 9B agentic model fine-tuned from Qwen3.5-9B, on Hugging Face under the MIT license. The model covers coding, general agent tasks, visual coding, and cybersecurity.
China context
- Original name
- 小米 MiMo
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download and fine-tune the model under the MIT license, using the provided tokenizer and chat template with SGLang.
- For investors
- Xiaomi's release of a competitive open-weight agentic model signals its commitment to the open-source AI ecosystem, which may influence the competitive landscape for agentic AI startups.
Xiaomi MiMo released MiMo-V2.6-Distill-Qwen-9B on Hugging Face. The model is a 9B agentic model developed through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. It covers coding, general-purpose agent tasks, visual coding, and cybersecurity. The release is intended as a starting point for open research in agentic reinforcement learning. The model is available under the MIT license and uses the transformers library.
The model is a 9B parameter agentic model fine-tuned from Qwen3.5-9B. It uses a weighted SFT data mixture of 77.4B total tokens, with 27.2B loss-bearing tokens across code, cyber, general, and visual domains. The checkpoint includes its tokenizer and MiMo v2.6 chat template, and requires a recent SGLang build with Qwen3.5 support for text generation.
Developers outside China gain immediate access to a 9B open-weight agentic model under the MIT license, enabling them to build and fine-tune agent applications without licensing costs. Xiaomi's release of a competitive agentic model on Hugging Face pressures other open-weight providers to match its capabilities in coding and cybersecurity benchmarks.
The MIT license allows commercial use, modification, and redistribution, making the model suitable for enterprise and startup applications in coding, cybersecurity, and general agent tasks. The model's strong performance on internal benchmarks suggests cost-effective deployment in agentic workflows.
Watch for community fine-tunes and RL training runs based on this SFT checkpoint, as the model card explicitly positions it as a starting point for agentic reinforcement learning research. Independent evaluations of the reported benchmark scores will clarify the model's real-world performance.