MiniMax M2
MiniMaxAI · released 2025-10-23 · other license
MiniMax-M2 is a compact 230B-total/10B-active open-weight MoE (Oct 2025, MIT) for coding and agentic work. Artificial Analysis rates it above-average for its size (index 36), and it ranked #1 open-weight at launch on the v3.0 index.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 228.7B total · MoE, — active |
| Architecture | minimax_m2 |
| Context window | 205K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 78% |
|---|---|
| SWE-bench Verified | 69.4% |
| AIME | 78% |
| MMLU-Pro | 82% |
| BFCL v3 (tool use) | — |
| Composite score | 6.38 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 134.1 GB | 132.6 GB | — | 205K |
Strengths & weaknesses
Strengths: Highly efficient agentic/coding MoE (10B active of 230B); Top-tier open-weight coding (LiveCodeBench 83) at low cost; MIT license
Weaknesses: Low standalone reasoning without tools (HLE 12.5); Text-only, somewhat verbose