XiaomiMiMo/MiMo-V2.6-Pro-MOPD
One-line: MOPD (Multi-teacher On-Policy Distillation) upgrade of MiMo-V2.6-Pro-RL — flagship 1.02T/42B-active omnimodal with tool-call-repetition fix, MIT, 1M ctx.
⚠ large — 1.02T total params; datacenter-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) | 71.9 | lab (V2.6 report, RL ckpt — holds steady) |
| ProgramBench | 26.5 | lab (V2.6 report, RL ckpt) |
| MiMo Code Bench (in-house) | 63.2 | lab (V2.6 report, RL ckpt) |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
| Toolathlon-Verified | 76.9 | lab (V2.6 report, RL ckpt) |
| AutomationBench v1.0.6 | 53.1 | lab (V2.6 report, RL ckpt) |
| Terminal Bench 2.1 | 89.9 | lab (V2.6 report, RL ckpt) |
| Agents’ Last Exam | 31.6 | lab (V2.6 report, RL ckpt) |
| JobBench | 62.0 | lab (V2.6 report, RL ckpt) |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| AA Intelligence Index | 46 (#1/114 open-weight) | Artificial Analysis (independent, on Pro-RL) |
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH | — | — |
| AIME 2026 | — | — |
Efficiency
- Active params: 42B (of 1.02T total; 384 routed experts, 8 active; ~70 layers hybrid full+SWA, no shared experts, MTP)
- Context length: 1M tokens
- VRAM (fp16 rough): ~2TB fp16 — NOT locally runnable; on-disk weights ~1TB (fp8/bf16 mix), needs multi-node expert-parallel or API.
What makes it notable
The flagship V2.6 checkpoint now has a production-grade agentic patch: Pro-RL’s exact within-turn tool-call repetition rate (0.54% under OpenCode, 0.10% under Claude Code) drops substantially after MOPD across every harness and context-length bucket, per Xiaomi’s published heatmaps, while the broader suite holds steady — the #1 open-weight AA Intelligence Index position (46) carries over from Pro-RL. The MOPD2 recipe (mixRL + SFT teacher fusion, teacher-prefix and SFT-prefix on-policy distillation) is a genuinely novel post-training algorithm that generalizes beyond the repetition fix to domains with unreliable verification (long-horizon game dev, science, embodied). For the user’s local/agent use case this remains a datacenter heavyweight — the 15B-active Flash-MOPD sibling is the realistic self-hosted pick; both are worth running via API given the agent-harness reliability fix.