MiniMax offers several text models tailored for different applications, with a focus on programming and office productivity. The standout model, MiniMax-M2.5, sets new benchmarks in various scenarios, achieving high performance with an output speed of about 60 tokens per second (TPS). For those needing even faster performance, the MiniMax-M2.5-highspeed variant doubles that to roughly 100 TPS. All models operate with a context window of 204,800 tokens, allowing for extensive input and output capabilities. M2.5 is particularly notable for its programming features, having been trained on over 10 programming languages like Python, Java, and Rust. It approaches coding by breaking down tasks from an architectural perspective, enhancing planning and execution. In office productivity, M2.5 integrates expert knowledge in finance and law, producing high-quality outputs in applications like Word and Excel, which can be crucial for tasks such as financial modeling. The model's efficiency is marked by a significant reduction in task completion time, with an average drop from 31.3 minutes to 22.8 minutes on the SWE-Bench, representing a 37% speedup. Cost efficiency is another key aspect; M2.5 can operate continuously at 100 TPS for about $1 per hour, making it viable for long-term deployments. This pricing supports the sustainable use of complex agents, enabling organizations to leverage these models without prohibitive costs. Users can access MiniMax models through various APIs, including those from Anthropic and OpenAI, and can integrate M2.5 into coding tools like Claude Code and Cursor. For support, users can reach out via email or GitHub.