MiniMax: MiniMax M2-her
Model Type
Proprietary Model
API access only
Recommended Use Cases
Text Generation
Try MiniMax M2-her
Overview
MiniMax M2-her is MiniMax's role-play specialist model designed for multi-character interactions, immersive long-horizon storytelling, and emotional dialogue, released in January 2026. The model represents three years of role-play optimization distilled into a production-grade model. It's built on a framework of Worlds (setting fidelity), Stories (narrative progression), and User Preferences (implicit intent recognition). The model powers MiniMax's Talkie AI companion app and is available via API for developers.
Key Features
- High-fidelity world experience: Maintains strict coherence with complex settings and character identities
- Dynamic story progression: Actively drives plot forward with vivid prose, avoiding repetitive loops
- Intuitive preference alignment: Detects unspoken user expectations and adapts without explicit instruction
- Multi-character coordination: Handles ensemble casts with distinct voices and relationships
- Long-horizon stability: Maintains quality across 100+ turn conversations
- Response length control: Keeps outputs within optimal range even in extended sessions
Design Philosophy
MiniMax defines deep role-play as navigating specific coordinates across three dimensions:
- Worlds: Preserving the distinct soul of each setting without regression to average
- Stories: Sustaining narrative vitality with proper cadence of conflict and emotional processing
- User Preferences: Decoding implicit intent for pacing and tone alignment
Capabilities
- Character customization with distinct personalities
- Emotional expression and relationship building
- Multi-turn dialogue with narrative progression
- Multi-character scenes with voice separation
- Narrative bridging for scene transitions
- User intent recognition without explicit instruction
Role in Series
MiniMax text models serve different use cases:
- MiniMax M2: Agentic workhorse, open weights, best cost-efficiency
- MiniMax M2.1: Multi-language coding specialist
- MiniMax M2.1-lightning: Same as M2.1 with faster inference
- MiniMax M2-her: Role-play and conversation specialist (this model)
- MiniMax M1: Reasoning model with extended thinking