Xiaomi · multimodal

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

MiMo-V2.6-Distill-Qwen-9B is Xiaomi's 9B supervised fine-tuning checkpoint built from Qwen3.5-9B using MiMo-generated agent data. It targets coding, tool use and visual coding; it is separate from the hosted MiMo Pro and Flash models.

Downloadable checkpoint · MIT (declared in official model card)

Official sources reviewed 2026-10-04. This page documents access and specifications; it does not report a model benchmark.

Task and deployment contract

Download the SFT checkpoint and use a recent SGLang build with Qwen3.5 support and the MiMo reasoning parser.

Xiaomi describes this checkpoint as a starting point for agentic reinforcement-learning research. It is an SFT distillation model, not the Pro/Flash RL or MOPD checkpoint.

The official text-generation example uses SGLang and explicitly enables thinking through the chat template. Preserve the bundled tokenizer and MiMo v2.6 template when serving.

The model card reports coding, general agent, visual coding and cybersecurity evaluations. Those are developer-reported results; this page does not import them as independent benchmark measurements.

The official model card declares MIT. No separate root LICENSE file was found in this reviewed repository revision; review the model card, base model and dependencies before distribution.

No current hosted price for this exact 9B checkpoint has been verified. Do not reuse the MiMo-V2.6-Pro or Flash API token rates for this downloadable model.

Billing and limits

No current provider rate for this exact configuration is recorded here. Open weights still require compute and operations; an unlisted rate does not mean a hosted API is free.

The checkpoint text configuration declares max_position_embeddings=262144. This architecture setting is not a tested service limit or a guarantee of accurate full-window retrieval.

Task and deployment contract

Billing and limits

Official access and sources