MiniMax releases M2.5 model with 80.2% SWE-Bench score

MiniMax

MiniMax released M2.5, achieving 80.2% accuracy on SWE-Bench Verified coding benchmarks. The M2.5-Lightning variant operates at 100 tokens per second, approximately 37% faster end-to-end than M2.1, while maintaining the cost advantage of roughly $1 per hour of inference. The model is available under MIT license on Hugging Face.

• SWE-Bench Verified: 80.2%; BrowseComp: 76.3%; BFCL tool use: 76.8%
• M2.5-Lightning achieves 100 tokens/sec, 37% faster than M2.1
• Priced at 1/10th to 1/20th the cost of Anthropic, Google and OpenAI offerings
• MIT-licensed; freely available on Hugging Face and platform.minimax.io