Chinese AI startup MiniMax has released M2.5, a new large language model that the company says competes with top-tier models from Anthropic and Google while costing significantly less to run.
MiniMax offers two versions through its API. M2.5-Lightning runs at 100 tokens per second and costs $0.30 per million input tokens and $2.40 per million output tokens. The standard M2.5 runs at half the speed for half the price. According to MiniMax, this makes the model between ten and twenty times cheaper than Claude Opus 4.6 based on output pricing.
The model uses a Mixture of Experts architecture with 230 billion total parameters, but activates only 10 billion at a time. MiniMax trained it using a proprietary reinforcement learning framework called Forge across hundreds of thousands of real-world environments, including coding tasks, document editing, and web search.
On the SWE-Bench Verified coding benchmark, M2.5 scored 80.2%. On BrowseComp, a test of web research ability, it scored 76.3%. MiniMax reports that M2.5 completed the SWE-Bench evaluation 37% faster than its predecessor.
The company states that M2.5 now handles 30% of internal tasks at MiniMax and generates 80% of newly committed code. The model has been released as open source on Hugging Face under a modified MIT License.
Sources: MiniMax, VentureBeat
