MiniMax releases M2-her role-play model
MiniMax has introduced MiniMax-M2-her, a role-play model built on three years of user data from its Talkie and Xingye apps, targeting character consistency and narrative coherence across long conversations.
MiniMax detailed MiniMax-M2-her in a research post. The model is built for production role-play and draws on what the company calls its third year of optimizing role-play in its Talkie and Xingye companion apps.
The post explains the design using user behavior the company has observed. Use of the "Regenerate" button clusters at narrative turning points such as confessions, which MiniMax reads as users chasing "narrative precision" rather than a simple pass/fail judgment of a reply. Popularity among characters does not follow a typical power-law curve: niche characters keep loyal, high-retention audiences, so a model tuned to the "average" user risks flattening them. MiniMax also reports a significant drop in conversation turns after turn 20.
From these observations the company defines role-play as navigating "{World} × {Stories}, conditioned on {User Preferences}." It names three challenges: keeping each world's distinct voice, avoiding repetitive loops and coherence drift over long sessions, and inferring pacing preferences that users don't state.
MiniMax says that before mid-2024 it relied on online A/B testing using metrics such as lifetime, session duration and conversation turns. Feedback cycles could take a week or more. Users also pushed back when a model swap changed a character's established voice. That led the company to build offline evaluation meant to approximate online metrics.
The material available here is cut off partway through the post. According to the signal summary, the full write-up covers a self-built 100-turn benchmark, an agentic data-synthesis pipeline, and online RLHF trained on user signals such as regenerations and session length. The model's reported "#1 overall" ranking comes from MiniMax's own benchmark. No independent results were provided.