MiMo V2.6 Distill Qwen 9B

MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model from Xiaomi MiMo, created by supervised fine-tuning Qwen/Qwen3.5-9B on MiMo-generated data. It is released under the MIT license and covers coding, general-purpose agent tasks, visual coding, and cybersecurity. The model card describes it as an SFT checkpoint intended as a starting point for open research in agentic reinforcement learning. It supports image-text-to-text input and includes safetensors and code tags. Reported evaluation results show improvements over Qwen3.5-9B on benchmarks such as SWE Verified, SWE Pro, Terminal Bench 2.1, and Toolathlon-Verified. The training data mixture contains 77.4B total tokens, of which 27.2B are loss-bearing. The checkpoint includes its tokenizer and MiMo v2.6 chat template, and a quickstart example is provided using SGLang with an OpenAI-compatible endpoint; thinking can be enabled via chat template kwargs.