MiniMax: MiniMax M2.1

MiniMax M2.1, a lightweight 10-billion parameter model, demonstrates a strong performance profile, particularly excelling in coding and general knowledge tasks. While its speed performance is moderate, ranking in the 25th percentile, and its pricing is also moderate at the 27th percentile, the model boasts exceptional reliability with a 99% success rate across benchmarks, indicating minimal technical failures. A key strength of M2.1 is its coding capability, achieving 95.0% accuracy in the Coding (Baseline) benchmark (92nd percentile) and leading multilingual coding performance with 49.4% on Multi-SWE-Bench and 72.5% on SWE-Bench Multilingual. It also shows high accuracy in General Knowledge (99.5%, 73rd percentile), Email Classification (99.0%, 81st percentile), and Reasoning (94.0%, 79th percentile). Notable weaknesses include its performance in Hallucinations (90.0% accuracy, 39th percentile) and Instruction Following (54.0% accuracy, 49th percentile), suggesting areas for improvement in handling uncertainty and complex multi-step directives. Despite these, M2.1's overall performance positions it as a versatile agent "brain" for development and general assistance, offering cleaner outputs and faster perceived response times than its predecessor.