MiniMax-M2.5

MiniMax-M2.5 was released in February 2026 and is positioned as a "native Agent production-level model". The goal is to solve the cost and performance bottlenecks in the large-scale implementation of AI Agent through breakthroughs in programming, tool calling and reasoning efficiency. M2.5 scored 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench. Its performance has significantly improved compared to M2.1. Its outstanding feature is that it exhibits "native Spec capabilities", that is, it proactively disassembles the architecture and plans functions before coding, which is closer to the working mode of a real architect. In terms of inference speed, M2.5 completes SWE-Bench evaluation 37% faster than M2.1, and the Lightning version supports output of more than 100 TPS. Cost is the core selling point of M2.5: it only costs about $1 to run continuously at 100 tokens/second for one hour, and it is positioned as "the first cutting-edge model where users do not need to worry about cost."

Context

204800 tokens

Max output

32768 tokens

Input

llm

Output

llm