MiniMax M2-her
MiniMax · 2026-01-27
MiniMax M2-her is MiniMax’s hosted text-dialogue model for role-playing and multi-turn conversations, with character customization, long-horizon interaction, and preference-aligned storytelling. MiniMax describes training it with agentic data synthesis followed by online preference learning using RLHF; the Her-series weights are not publicly released.
Benchmark scores
| Benchmark | Score |
|---|---|
| BuseyBench SVG | 2.2 |
| Longform Creative Writing | 34.8 |
| SnakeBench Average Apples | 0.8 |
| SnakeBench Best Apples | 3.0 |
| SnakeBench Rating | 15.5 |
| SnakeBench Total Apples | 10.0 |
| SnakeBench Win Rate | 46.2 |