XiaomiMiMo/MiMo-V2.6-Flash-MOPD
One-line: MOPD (Multi-teacher On-Policy Distillation) upgrade of MiMo-V2.6-Flash-RL that fixes tool-call repetition in agent harnesses — same 309B/15B-active, 1M-ctx, MIT omnimodal stack.
⚠ large — 309B total params; server-class hardware only (see Efficiency). ⚠ Lab-reported benchmarks below are the MiMo-V2.6 technical-report numbers measured on the RL checkpoint; the MOPD checkpoint’s card publishes no new eval tables, but Xiaomi’s blog (2026-09-27) states the broader benchmark suite “held steady” after MOPD. Released Sep 27, 2026 (weights on HF 04:00Z, API since Sep 25).
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 (V2.6 report, RL ckpt — holds steady) |
| ProgramBench | 26.0 | lab (V2.6 report, RL ckpt) |
| MiMo Code Bench (in-house) | 61.2 | lab (V2.6 report, RL ckpt) |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
| Toolathlon-Verified | 73.6 | lab (V2.6 report, RL ckpt) |
| AutomationBench v1.0.6 | 52.3 | lab (V2.6 report, RL ckpt) |
| Terminal Bench 2.1 | 87.6 | lab (V2.6 report, RL ckpt) |
| Agents’ Last Exam | 27.6 | lab (V2.6 report, RL ckpt) |
| JobBench | 61.2 | lab (V2.6 report, RL ckpt) |
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; 48 layers hybrid full+SWA, MTP speculative decoder)
- Context length: 1M tokens
- VRAM (fp16 rough): ~618GB fp16 — NOT consumer-runnable; on-disk ~311GB (fp8/bf16 mix), needs multi-GPU + quant.
What makes it notable
The MOPD checkpoint is the agentic-reliability patch for the MiMo-V2.6-Flash family: RL scaling surfaced a tool-call-repetition failure mode (e.g. 1.02% exact within-turn repetition rate under OpenCode on Flash-RL) that standard correctness optimization missed. Xiaomi’s fix fuses mixRL + SFT teachers via on-policy distillation (standard MOPD, teacher-prefix OPD, SFT-prefix OPD) in a lightweight run (~$90k total, ~4% of the MixRL restart cost) that folds a repetition-specialized teacher in without moving the broader benchmark suite. For agent-loop users this targets a real operational failure — repeated identical tool calls burning context and wall time — rather than a headline eval. Same 309B/15B-active cost profile as Flash-RL: server-class only; the 9B Distill sibling remains the local entry point.