Agentic workflows (tool use + environment interaction) often need a practical SFT seed that already encodes code reasoning, UI/terminal manipulation and multimodal perception. MiMo-V2.6-Distill-Qwen-9B fills that role: a 9B supervised-fine-tuned checkpoint distilled from Qwen3.5-9B and trained on a MiMo-generated mixture to make downstream RL training and agent-harness experimentation faster and more sample-efficient.
Key Capabilities
- Multimodal agentic SFT: fine-tuned on a mixture covering code, cyber tasks, general automation and visual coding, so it provides sensible initial behavior for tool use and environment interaction without full RL tuning.
- Measurable SFT improvements: shows notable gains over the Qwen3.5-9B SFT baseline across internal benchmarks (code, terminal automation, cyber, visual coding), which means fewer RL iterations are likely needed to reach acceptable agent performance.
- Research-oriented packaging: includes tokenizer, MiMo v2.6 chat template and instructions for serving via SGLang, so teams can quickly integrate it into RL pipelines, harnesses and evaluation environments.
- Compact/portable: at 9B parameters and provided in safetensors, it is more tractable for research labs than Pro-scale models while still carrying agentic priors useful for long-horizon and tool-using tasks.
Who it's for and tradeoffs
Great fit if you are a researcher or engineer who wants a ready SFT checkpoint to bootstrap agentic RL training, compare reward designs or iterate on harness components without starting from a base LLM. It is also suitable for prototyping code-centric and visual coding agents that need a small-to-medium footprint model. Look elsewhere if you need a production-grade RL-optimized policy (this is an SFT checkpoint, not the final RL-trained Pro model), absolute state-of-the-art closed-source performance, or strict provenance for every training datum. Running at reasonable latency still requires GPU resources and integration with SGLang or compatible serving stacks.
Where it fits
Positioned between base LLMs (e.g., raw Qwen3.5-9B) and full RL-finetuned MiMo-V2.6-Pro agents: use this checkpoint to shorten RL training cycles, to evaluate harness and reward changes, or to reproduce MiMo's SFT-to-RL workflow without needing Pro-scale compute.
