XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
One-line: The locally-runnable member of the MiMo-V2.6 family — 9B agentic SFT of Qwen3.5-9B on MiMo-generated data, SWE-bench Verified 61.1, MIT, vision+text.
⚠ Vision-language model (image-text-to-text) — vision encoder present, but it has real published coding/agentic evals, so it qualifies for the coding/agent tracker. ⚠ Lab-reported numbers (HF card, from the MiMo-V2.6 technical report). Avg@3 for code/cyber/visual, avg@1 for general.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | — |
| HumanEval+ | — | — |
| MBPP | — | — |
| LiveCodeBench | — | — |
| Aider polyglot | — | — |
| SWE-bench Verified (avg@3) | 61.1 | lab |
| SWE-bench Multilingual | — | — |
| SWE Pro (avg@3) | 44.6 | lab |
| MiMo Code (mini, avg@3, in-house) | 51.6 | lab |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
| Toolathlon-Verified (avg@1) | 35.2 | lab |
| AutomationBench v1.0.6 (avg@1) | 30.3 | lab |
| Terminal Bench 2.1 (avg@1) | 37.1 | lab |
| JobBench (avg@1) | 18.3 | lab |
| MiMo Cyber (mini, in-house) | 31.3 | lab |
| MiMo Visual Coding (mini, in-house) | 64.0 | lab |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH | — | — |
| AIME 2026 | — | — |
Efficiency
- Active params: 9B dense (32 layers, hidden 4096) + 1152-hidden vision encoder; distilled from MiMo-V2.6 onto Qwen3.5-9B
- Context length: 256K tokens (max_position_embeddings 262144)
- VRAM (fp16 rough): ~18GB fp16 — runs on a 24GB GPU at fp16, ~6GB at Q4; on-disk ~9.4GB. The practical local agent model of the family.
- Serving: SGLang with Qwen3.5 support +
--reasoning-parser mimo; chat template supportsenable_thinking.
What makes it notable
SFT checkpoint on MiMo-generated data (77.4B tokens: code 23.2B, cyber 11B, general 22B, visual 21.2B) as the open starting point for agentic-RL research. vs its Qwen3.5-9B base: SWE-bench Verified 60.0→61.1, SWE Pro 32.0→44.6 (+12.6pp), AutomationBench 5.0→30.3, Terminal Bench 2.1 27.0→37.1 — big agentic gains from data alone. Vision + 256K ctx + MIT at 9B makes it the natural local candidate for tool-using agents that also need screenshots/DOM images, if the SGLang/transformers support is stable (custom tokenizer + chat template; verify before betting an agent loop on it).