MiniMax releases MiniMax-M2.5 and M2.5-Lightning

Background

MiniMax had already built coding and agent models. In February 2026 it released M2.5 in two versions: a standard model, and M2.5-Lightning, a faster variant aimed at the same kind of work ^[1].

What happened

Both models were released in February 2026. The weights of M2.5 were posted to Hugging Face on 12 February 2026 ^[2][3]. Hugging Face is a platform where researchers and developers share models and datasets, and posting the weights there lets outside users download the model.

Key details

The Lightning version runs at a steady 100 tokens per second, and the standard version at 50. A token is the unit of text a model reads and writes, roughly a word or part of a word ^[1]. MiniMax said that running the model for an hour costs $1 at 100 tokens per second and $0.30 at 50 tokens per second ^[1]. Lightning is priced at $0.3 per million input tokens and $2.4 per million output tokens ^[1].

MiniMax reported 80.2% on SWE-bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp, the last with context management. Context management is a way of keeping the most relevant information within the model's working window during a long task ^[1].

The models were trained with Forge, an in-house reinforcement-learning framework built for agents. MiniMax says Forge speeds up training about 40 times. Training also used the CISPO algorithm and a process reward mechanism, which scores intermediate steps rather than only the final result ^[1].

Impact

The release gave users two speed and price options. The Lightning version produces text at twice the rate of the standard version, and running it for an hour costs $1 against $0.30 ^[1]. The release extended the M2 line for coding and agent work. The next model in the line, M2.7, followed in March 2026 ^[2].