MiniMax-M3: benchmark results
MiniMax M3 model for reasoning and agentic tasks. Provider: MiniMax. Released 2026-06-01. Access: Open.
Aggregate Rating 1734 ± 4, rank #136 of 1656 rated models, from 1076 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ArkEval (Official-Knowledge RAG) - Compile@1 | 51.79 | Compile@1 (%): share of the 502 instances whose patch comple | 100 |
| ClashBench - Deliberate Interference Rate (OpenCode) | 46.1 | DIR (%) | 100 |
| DrawAI-Bench - Editability - Table | 90.7 | Editability score (0-100; rules and VLM rubric, strongest ob | 100 |
| Evo-Bench - Office | 41.7 | Held-out score (%; the fixed DeepSeek-V4-Flash policy runs t | 100 |
| FinCUABuildBench (Tool-Augmented Prompting) | 17.5 | Qualified task rate (%; share of the 576 preregistered const | 100 |
| Interpretive Canons (BVerfG) - Argument Gate | 74.7 | Positive-class F1 (%; argument gate (argument): whether a ca | 100 |
| Interpretive Canons (BVerfG) - Not Abstract | 80.6 | Positive-class F1 (%; not abstract (nicht_abstrakt): whether | 100 |
| Jais 3 Model Cards - AA-LCR (Arabic) | 80 | Accuracy (%) | 100 |
| KC-Bench - Region | 67.61 | Success rate (%; world-knowledge conflict: 71 region tasks i | 100 |
| LLM Stats (Claw-Eval) | 74.5 | Score (%) | 100 |
| LLM Stats (OmniDocBench 1.5) | 91.6 | Score (%) | 100 |
| AA IFBench | 82.86 | Accuracy (%) | 99.6 |
Interactive version: theaggregate.ai/model?slug=minimax-m3 · How It Works · Data refreshed daily, snapshot 2026-10-20.