What is MiniMax-M2-her
MiniMax-M2-her is a role-playing model launched by MiniMax specifically for deep AI companionship scenarios, and is the underlying model servingXingye/Talkie . Through three core capabilities—unique world experience, rhythmic story progression, and precise understanding of user preferences—the model addresses pain points in long conversations such as character breakdown and plot repetition. Based on the innovative Role-Play Bench evaluation system and Agentic data synthesis technology, the model ranks first overall in 100-turn long-horizon dialogue evaluations and is now available for API access.
Main features of MiniMax-M2-her
- World-building: The model accurately understands and maintains complex worldview settings, supports collaborative role-playing with multiple characters and narration, and ensures that character relationships and spatial logic always remain clear and consistent.
- Story progression: The model can actively drive the plot forward, avoiding mechanical repetition and formulaic narration, so that the story maintains a sense of breathing and vitality across long and short rhythms.
- Preference perception: It captures unspoken expectations from the details of user interactions, dynamically adapts to different narrative rhythms and interaction styles, and delivers personalized experiences for each individual.
- Long-horizon stability: Across hundreds of turns of dialogue, it continuously maintains persona consistency, logical coherence, and controllable response length, solving the problem of quality degradation in long conversations.
- Safe interaction: Within a compliance framework, it flexibly grasps boundaries, avoids excessively rejecting users' reasonable interaction needs, and balances safety and immersion.
Technical principles of MiniMax-M2-her
- Role-Play Bench evaluation system: Addressing the characteristic that role-playing scenarios have no standard answers, it proposes an evaluation framework based on "Misalignment." Through a Model-on-Model self-play mechanism, it generates multi-turn dialogue trajectories and automatically detects the model's misaligned behaviors from three dimensions—Worlds (worldview consistency), Stories (narrative quality), and User Preferences (interaction appropriateness)—achieving rapid offline alignment with real user experience.
- Agentic Data Synthesis: It builds a dialogue pipeline with dual expert models, where expert models playing the user and NPC respectively generate candidate dialogue turns. After multi-dimensional scoring by a Reward Model, the Best-of-N strategy selects the optimal response, and an LLM-as-judge checkpoint mechanism is triggered periodically to correct logical errors, referential confusion, and content repetition. At the same time, a planning agent is introduced to dynamically assess the dialogue state and suggest plot directions, combining scene diversification, Prompt expansion, a style expert library, and dynamic turn allocation strategies to ensure data diversity and quality.
- Online Preference Learning: It collects explicit user feedback (regenerate, like) and implicit signals (dwell time) in the product environment. After hierarchical sampling, causal inference, and outlier filtering to remove noise, RLHF is used to train the model to perceive contextualized preferences. During training, output diversity is continuously monitored and early stopping is applied before mode collapse, forming a positive cycle of "model deployment-user interaction-signal collection-iterative training" and continuously raising the ceiling of user preference alignment.
Project address of MiniMax-M2-her
- Project official website：https://www.minimaxi.com/news/minimax-m2-her-%E6%8A%80%E6%9C%AF%E6%B7%B1%E5%BA%A6%E8%A7%A3%E6%9E%90
- API access address：https://platform.minimaxi.com/docs/api-reference/text-chat
Application scenarios of MiniMax-M2-her
- AI emotional companionship: It provides deep, long-horizon, personalized role-playing experiences for virtual character chat products such as Xingye/Talkie, helping users build stable emotional connections with AI.
- Interactive narrative games: Dynamically advance plot branches, support multi-character ensemble narratives, and deliver an open-world-style story exploration experience for text adventure, dating simulation, and mystery puzzle games.
- Virtual IP Operations: Precisely maintain the persona and worldview of anime, game characters, or celebrity virtual avatars, giving fans an immersive experience like conversing with the real character.
- Creative Writing Assistance: Understand complex settings and proactively develop plots, serving as an intelligent collaborator for authors to expand the narrative possibilities of novel co-creation, script development, and worldbuilding.
- Language Learning: Simulate real interaction scenarios, dynamically adjust dialogue difficulty and pace according to the learner's level, and enable contextualized spoken dialogue and role-playing-based teaching.