MiniMax-M2.5

What it is

MiniMax-M2.5 is a large language model released by MiniMax in February 2026 with open weights. The checkpoint is listed at 229B parameters and is published under a modified MIT license, so the weights can be downloaded and run outside the company's own systems ^[1]. The weights were posted to Hugging Face on 12 February 2026, the same day the model was released together with a faster variant called MiniMax-M2.5-Lightning ^[2][3]. The model is built for agent work: writing and repairing code, browsing, and office tasks that take many steps rather than a single answer.

How it was trained

MiniMax trained the model with reinforcement learning across hundreds of thousands of complex real-world environments ^[1]. The training stack is built around Forge, an in-house agent-native reinforcement learning framework that MiniMax says speeds training up by roughly 40 times. Forge is used together with the CISPO algorithm and a process reward mechanism, which scores the intermediate steps of a task as well as its final answer ^[1].

Two versions, and what they cost

The two versions differ in speed, not capability. MiniMax says Lightning is identical in capability to M2.5, and that both support caching ^[1]. Lightning runs at a steady 100 tokens per second, while the standard version runs at 50 ^[1]. MiniMax priced running the model by the hour: about $1 per hour at 100 tokens per second, and $0.30 per hour at 50 tokens per second ^[1]. For the Lightning variant specifically, the listed API price is $0.3 per million input tokens and $2.4 per million output tokens ^[1]. MiniMax also said that the standard M2.5 costs half as much as the Lightning variant at 50 tokens per second ^[1].

Benchmark results

The headline numbers MiniMax reports are SWE-bench Verified 80.2%, Multi-SWE-Bench 51.3%, and BrowseComp 76.3% when the model is given context management ^[1]. On third-party scaffolds the margins are narrower: through the Droid scaffold M2.5 scored 79.7 on SWE-Bench Verified against 78.9 for Claude Opus 4.6, and through OpenCode it scored 76.1 against 75.9 ^[1]. The model card's evaluation metadata also lists SWE-Bench Pro at 55.4 ^[1].

Speed and token consumption were part of the reported results as well. MiniMax says M2.5 completes the SWE-Bench Verified evaluation 37% faster than the earlier M2.1 and matches the speed of Claude Opus 4.6 ^[1]. It averages 3.52M tokens per task on SWE-Bench Verified, against 3.72M for M2.1 ^[1].

Use inside MiniMax

MiniMax reported internal usage figures: about 30% of its internal tasks are completed autonomously by M2.5, and about 80% of newly committed code is produced by the model ^[1]. On its internal GDPval-MM office-work evaluation, the company says M2.5 has an average win rate of 59.0% ^[1]. These are MiniMax's own numbers rather than independent measurements.

Where it sits

In March 2026 MiniMax followed M2.5 with MiniMax-M2.7 ^[2].