MiniMax: MiniMax M2-her

MiniMax M2-her demonstrates moderate speed performance, ranking in the 27th percentile, and offers competitive pricing, placing it in the 52nd percentile across benchmarks. This dialogue-first model, designed for immersive roleplay and expressive multi-turn conversations, exhibits a mixed performance profile across various categories. Its primary strength lies in its competitive pricing. However, the model shows notable weaknesses in several key areas. Hallucinations (52.0% accuracy) and Instruction Following (27.0% accuracy) are particularly low, indicating challenges in acknowledging uncertainty and precisely adhering to complex directives. General Knowledge (65.5%), Ethics (61.0%), Email Classification (90.0%), Reasoning (50.0%), and Coding (25.0%) also fall within the lower percentiles, suggesting a need for improvement in factual recall, ethical judgment, classification accuracy, logical problem-solving, and programming proficiency. While its design emphasizes consistent tone and personality for conversational experiences, its current benchmark results suggest that its core capabilities in accuracy and instruction adherence require further development to fully leverage its intended applications in storytelling and companionship.