MiniMax-M2
230B-total, 10B-active open MoE built for coding and agents with interleaved thinking; scored 61 on the Artificial Analysis index, ranked first among open models.
Only 10B parameters active per token, aimed at fast agent loops (about 100 tokens/s, near double comparable models per MiniMax). It uses interleaved thinking, so reasoning inside think tags must be kept in history. Priced at about 8% of Claude Sonnet 4.5; modified-MIT licence.
- Date
- Monday, 27 October 2025
- Lab
- MiniMax
- Kind
- open-weights
- Access
- open weights
- Price
- $0.30 input / $1.20 output per M tokens, 2025-10
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 230B / 10B | company |
| SWE-bench Verified | 69.4 | company |
| Terminal-Bench | 46.3 | company |
| BrowseComp | 44.0 | company |
| Artificial Analysis Intelligence Index | 61 company-cited; top five overall at launch per MiniMax | company |
MiniMax release notes list 2025-10-27 (a tracker says 10-26); HF repo created 2025-10-22. Benchmark numbers are company-reported; the AA ranking is cited by MiniMax. Anthropic later alleged MiniMax ran 13M+ distillation exchanges against Claude (disclosed 2026-02-23); MiniMax's response not found.
Sources
- www.minimax.io/news/minimax-m2
- huggingface.co/MiniMaxAI/MiniMax-M2
- platform.minimax.io/docs/release-notes/models
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check