According to MiniMax, M2.5 was trained using reinforcement learning across hundreds of thousands of complex environments, enabling it to perform particularly well on long-running tasks while maintaining low token consumption.
Unlike models that rely solely on prompt-based planning, M2.5 has learned to optimize its actions through internal planning mechanisms. The model is also designed to handle Word, Excel, and PowerPoint files quickly and smoothly.
In benchmark comparisons published by MiniMax, M2.5 delivers strong programming performance: 80.2% on SWE-Bench Verified, 79.7% on the Droid platform, and 76.1% on OpenCode, in some cases narrowly outperforming Anthropic’s Claude Opus 4.6. The model also scores competitively in web search (BrowseComp: 76.3%), AI agent tool use (BFCL: 76.8%), and office productivity tasks such as Excel workflows, where it reportedly outperforms GPT-5.2, Gemini 3 Pro, and Claude in several categories.
Fast and cost-efficient
MiniMax offers two variants of the model. M2.5-Lightning delivers throughput of 100 tokens per second, roughly twice as fast as other leading models. Continuous operation for one hour reportedly costs just $1. MiniMax claims the model is 10–20× cheaper than comparable offerings from Anthropic, Google, and OpenAI, with four M2.5 instances able to run for an entire year for $10,000.
The model weights are freely available on Hugging Face and GitHub. A detailed technical report is published at minimax.io/news/minimax-m25, while API access and a dedicated coding plan are offered via platform.minimax.io.
Beyond language models, MiniMax is also active in video AI. In June 2025, the company released Hailuo 02, the second generation of its video model. Since September 2025, MiniMax has faced a lawsuit from Disney, Universal, and Warner Bros over alleged copyright infringement related to this model. Despite the legal dispute, the company completed a successful IPO in early January 2026, with its stock price doubling on the first trading day. MiniMax was founded in 2021 by former SenseTime employees.
ES
EN