MiniMax-M3 (Thinking): benchmark results
MiniMax's open-weight MiniMax M3 (June 2026), a native multimodal MoE model with about 428B total and 23B active parameters, using MiniMax Sparse Attention for a 1-million-token context. Provider: MiniMax. Released 2026-06-01. Access: Open.
Aggregate Rating 1687 ± 18, rank #401 of 2170 rated models, from 85 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Nejumi 4 - Toxicity - Fairness | 99.44 | Criteria met (%) | 94.1 |
| Nejumi 4 - BFCL - Multi-Turn | 61.11 | Accuracy (%) | 93.1 |
| Nejumi 4 - MT-Bench (Japanese) - Roleplay | 99 | Judge rating (1-10, x10) | 92 |
| Nejumi 4 - jaster (2-shot) - JSeM | 82 | Exact match (%) | 91.7 |
| Nejumi 4 - JBBQ (2-shot) - Accuracy | 94 | Accuracy (%) | 90.3 |
| Nejumi 4 - jaster (2-shot) - JaMP | 80 | Exact match (%) | 88.5 |
| Nejumi 4 - jaster (2-shot) - JCoLA (out-of-domain) | 87 | Exact match (%) | 87.5 |
| Nejumi 4 - MT-Bench (Japanese) - Writing | 98 | Judge rating (1-10, x10) | 84.4 |
| Nejumi 4 - jaster (0-shot) - JCoLA (in-domain) | 79 | Exact match (%) | 84.4 |
| Nejumi 4 - jaster (2-shot) - JNLI | 90 | Exact match (%) | 84.4 |
| Nejumi 4 - MT-Bench (Japanese) - Extraction | 96.5 | Judge rating (1-10, x10) | 83.3 |
| Nejumi 4 - HLE (Japanese) - Accuracy | 36.08 | Accuracy (%) | 83 |
Interactive version: theaggregate.ai/model?slug=minimax-m3-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.