MiniMax M2.5 is a Mixture-of-Experts model designed to deliver frontier-tier performance at a fraction of the typical compute cost. With 230 billion total parameters but only 10 billion activated during inference, the architecture keeps most of the model idle on any given pass, which is what makes the pricing viable. The model builds on the coding expertise of its predecessor M2.1 and extends into general office work, reaching fluency in generating and operating Word, Excel, and PowerPoint files, context switching between diverse software environments, and working across different agent and human teams. This full-stack agentic capability means M2.5 is not just a code model — it is positioned as a productivity workhorse that can handle tool calling, web search, and office workflows in a single session.
The model was trained using MiniMax's proprietary Forge reinforcement learning framework, which scales agent training across 200,000+ real-world environments including code repositories, web browsers, and office applications. The Forge framework uses a CISPO algorithm and achieves a 40x training speedup compared to earlier approaches. M2.5 scores 80.2% on SWE-Bench Verified — placing it within 0.6 percentage points of Claude Opus 4.6 — while also reaching 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp. Both variants ship as open weights on Hugging Face under a modified MIT license, giving teams the freedom to self-host, fine-tune, or integrate the model directly into their pipelines. The Lightning variant doubles the standard speed to around 100 tokens per second, making it practical for real-time coding agents and high-throughput production workloads.
