MiMo-V2.6 9B arrives in GGUF for local agents

Xiaomi’s MiMo team has released MiMo-V2.6-Distill-Qwen-9B, a compact model tuned for coding, tool use, terminal work, cybersecurity tasks, and image-based programming. Community conversions now make it practical to run locally: Bartowski’s GGUF repo ranges from roughly 3.5 GB low-bit files to 17.9 GB BF16 weights, with the recommended Q4_K_M version at 5.84 GB.

A 5.84 GB model can fit its weights in many 8 GB GPUs, although the context cache, runtime buffers, and vision projector also consume memory. Developers with tighter limits can offload some layers to system RAM, while 12 GB or larger GPUs leave more room for longer contexts and image input.

A smaller agent trained by MiMo

MiMo-V2.6-Distill-Qwen-9B starts from Qwen3.5-9B and uses supervised fine-tuning on examples generated by Xiaomi’s larger MiMo models. In this context, “agentic” means the model was trained for multi-step work involving tools, terminals, code repositories, browsers, and visual interfaces.

The checkpoint accompanies Xiaomi’s larger MiMo-V2.6 Pro and Flash models, which process several input types. The broader release described in the MiMo release notes includes model weights, a technical report, more than 7,000 reinforcement-learning task environments, an end-to-end RL framework, and smaller evaluation harnesses.

Reported gains over Qwen3.5-9B

Xiaomi says the supervised fine-tuning mixture contained 77.4 billion processed tokens, including 27.2 billion loss-bearing tokens. Loss-bearing tokens are the target outputs used to update the model, while the larger total also includes prompt and context tokens. The mixture covered code, general agent work, visual tasks, and cybersecurity.