MiniMax M2.5 is a new frontier‑class large language model from MiniMax that targets high‑end coding, agents and office workloads while keeping costs relatively low and speed very high. Built on a roughly 229B‑parameter Mixture of Experts architecture (MoE) with a long context window of around 200k tokens, it focuses on efficient task decomposition and fast tool‑using behavior rather than just raw size. In coding benchmarks like SWE‑Bench Verified it reaches about 80% and outperforms the previous M2.1 generation by roughly a third in speed, landing in the same performance band as top commercial models such as Claude Opus and recent GPT‑series systems for code and agentic workflows.
The “Lightning” serving variant reaches about 100 tokens per second and is priced at around 0.3 USD per million input tokens with relatively low output pricing, which makes it attractive for always‑on agents, search pipelines and large‑scale automation. Compared with competitors, M2.5 usually trades a bit of general reasoning and multimodal breadth for strong coding, efficient tool calling and good price‑to‑performance, so it fits best where execution speed, cost and integration into complex workflows matter more than being the single most capable generalist model.
