XiaomiMiMo/MiMo-V2.6-Flash-RL
One-line: Cost-optimized V2.6 sibling — 309B total / 15B active, ~DeepSeek-V4.1-Flash-class agentic scores, native omnimodal, 1M ctx, MIT.
⚠ large — 309B total params; server-class hardware only (see Efficiency). ⚠ Lab-reported numbers (Xiaomi HF card + blog); see Pro-RL note for the benchmaxxing controversy context. Released Sep 21, 2026.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | — |
| HumanEval+ | — | — |
| MBPP | — | — |
| LiveCodeBench | — | — |
| Aider polyglot | — | — |
| SWE-bench Verified | — | not published |
| SWE-bench Multilingual | — | — |
| DeepSWE v1.1 (agentic SWE stand-in) | 67.9 | lab |
| ProgramBench | 26.0 | lab |
| MiMo Code Bench (in-house) | 61.2 | lab |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
| Toolathlon-Verified | 73.6 | lab |
| AutomationBench v1.0.6 | 52.3 | lab |
| Terminal Bench 2.1 | 87.6 | lab |
| Terminal Bench 4.0 | 28.8 | lab |
| Agents’ Last Exam | 27.6 | lab |
| JobBench | 61.2 | lab |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH | — | — |
| AIME 2026 | — | — |
Efficiency
- Active params: 15B (of 309B total; 256 routed experts, 8 active, top-k; 48 layers hybrid full+SWA, 3 MTP next-token-prediction layers)
- Context length: 1M tokens
- VRAM (fp16 rough): ~618GB fp16 — NOT consumer-runnable; on-disk ~311GB (fp8/bf16 mix), needs multi-GPU + quant. The “Flash” name means cheaper/smaller than Pro, not local.
What makes it notable
Xiaomi’s own appendix positions Flash as the value pick: DeepSWE v1.1 67.9 vs DeepSeek V4.1 Flash’s 74.2 and TB2.1 87.6 vs 90.6 (lab table — DeepSeek still edges it on those), but MiMo Flash leads on JobBench (61.2 vs 45.8) and MiMo Code Bench (61.2 vs 60.2). Huge improvement over MiMo-V2.5-Pro (DeepSWE 19.0 → 67.9). Same “You Only RL Once” mixed-reward recipe and 1M-ctx omnimodal stack as Pro, at lower active-param cost per token. Not a realistic local/agent-box model at 309B — the 9B Distill sibling is the local entry point.