M2-her

Description

MiniMax M2-her — conversational model built for role-play and multi-turn chat, with a 65,536-token context window and replies capped at 2,048 tokens (max_completion_tokens). Beyond system / user / assistant it accepts the vendor's extended message roles on the OpenAI-compatible chat endpoint: user_system (the user's persona), group (the conversation's name) and sample_message_user / sample_message_ai (example exchanges that steer the reply style). Every extended-role message must carry a name field; without it the vendor rejects the whole request with error 2013, while plain system/user/assistant messages need no name. Text only: no image input, and the vendor documents no tool calling or caching. Every call carries roughly 170 tokens of vendor-side persona overhead on top of the visible prompt, billed as input.