MiniMax-M2-her - MiniMax's AI role-playing model

What is MiniMax-M2-her?

MiniMax-M2-her is a role-playing model developed by MiniMax for deep AI companionship scenarios, serving as the underlying model for Hoshino/Talkie. The model addresses pain points such as character inconsistency and repetitive storylines in long conversations through three core capabilities: a unique world experience, rhythmic story progression, and accurate understanding of user preferences. Based on an innovative Role-Play Bench evaluation system and Agentic data synthesis technology, the model ranked first in overall performance across 100 rounds of long-term dialogue evaluation, and API access is now available.

MiniMax-M2-her main functions

• World BuildingThe model accurately understands and maintains complex world-building settings, supports multiple characters and narrators in collaborative performances, and ensures that character relationships and spatial logic remain clear and consistent at all times.
• Story progressionThe model can actively drive the plot forward, avoiding mechanical repetition and formulaic narratives, allowing the story to maintain a sense of breath and vitality in its varying lengths and rhythms.
• Preference perceptionCapture unspoken expectations from the details of user interaction, dynamically adapt to different narrative rhythms and interaction styles, and achieve a personalized experience for each user.
• Long-range stabilityMaintaining consistency in persona, logical coherence, and controllable response length throughout hundreds of rounds of dialogue solves the problem of quality decay in long conversations.
• Safe InteractionWithin the framework of compliance, we should flexibly grasp the boundaries, avoid excessively rejecting users' reasonable interaction needs, and balance security and immersion.

MiniMax-M2-her technical principles

• Role-Play Bench Evaluation SystemTo address the lack of standard answers in role-playing scenarios, this paper proposes an evaluation framework based on "misalignment." A multi-turn dialogue trajectory is generated through a Model-on-Model self-play mechanism, and inaccuracies in the model are automatically detected from three dimensions: Worlds (worldview consistency), Stories (narrative quality), and User Preferences (interaction appropriateness), enabling rapid offline alignment with real user experiences.
• Agentic Data SynthesisA dual-expert model dialogue pipeline is constructed, in which expert models playing the roles of user and NPC respectively generate candidate dialogue rounds. After multi-dimensional scoring by the Reward Model, the best response is selected through the Best-of-N strategy. The LLM-as-judge checkpoint mechanism is triggered periodically to correct logical errors, confusion of reference, and content repetition. At the same time, a planning agent is introduced to dynamically evaluate the dialogue state and suggest the plot direction. Combined with strategies such as scene fragmentation, Prompt expansion, style expert library, and dynamic round allocation, data diversity and quality are ensured.
• Online Preference LearningIn the product environment, explicit user feedback (repeated phrases, likes) and implicit signals (stay time) are collected. After removing noise through hierarchical sampling, causal inference and outlier filtering, the RLHF model is trained to perceive contextualized preferences. During the training process, the diversity of output is continuously monitored and stopped early before the pattern collapses, forming a positive cycle of "model deployment - user interaction - signal collection - iterative training" to continuously increase the upper limit of user preference alignment.

MiniMax-M2-her project address

• Project official websiteLink: 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-powered emotional companionshipProvides in-depth, long-term, and personalized role-playing experiences for virtual character chat products such as Hoshino/Talkie, helping users establish stable emotional connections with AI.
• Interactive narrative gamesDynamically advances plot branches, supports multi-character ensemble narratives, and enables an open-world story exploration experience for text adventure, romance simulation, and mystery puzzle games.
• Virtual IP OperationIt accurately maintains the character design and worldview of anime, game characters, or virtual celebrity avatars, giving fans an immersive experience as if they were interacting with real characters.
• Creative Writing AssistanceIt understands complex settings and actively develops the plot, acting as an intelligent collaborator with the author to expand the narrative possibilities of novel co-creation, script development, and world-building.
• Language learningSimulates real-life interactive scenarios, dynamically adjusts the difficulty and pace of dialogues according to learners' levels, and achieves contextualized oral dialogues and role-playing teaching.