MiniMax-M2.5 is a state-of-the-art large language model built for real-world productivity. Trained across a wide variety of complex digital work environments, M2.5 extends the coding strengths of M2.1 into broader office tasks—becoming fluent in creating and manipulating Word, Excel, and PowerPoint files, seamlessly switching context between different software tools, and collaborating across mixed agent and human teams. It scores 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp. Compared to earlier generations, M2.5 is also more token-efficient, having been trained to plan effectively in order to optimize both its actions and outputs.
MiniMax-M2.1 is a cutting-edge, lightweight large language model designed for coding, agentic workflows, and modern application development. With just 10 billion activated parameters, it offers a significant boost in real-world performance while ensuring low latency, high scalability, and cost-effectiveness. Compared to the previous version, M2.1 delivers more concise, clearer outputs and quicker response times. It excels in multilingual coding, achieving 49.4% on Multi-SWE-Bench and 72.5% on SWE-Bench Multilingual, making it an adaptable engine for IDEs, coding tools, and a wide range of assistant applications.
MiniMax-M2 is a compact, efficient language model with 10B active (230B total) parameters, optimized for coding and agentic workflows. It achieves near-frontier reasoning and tool use with low latency and deployment cost. The model excels in code generation, multi-file editing, compile-run-fix cycles, and automated test repair, showing strong results on SWE-Bench and Terminal-Bench. MiniMax-M2 performs well in agentic benchmarks like BrowseComp and GAIA, handling long-term planning, retrieval, and error recovery. With a small activation footprint, it delivers fast inference and high concurrency, making it ideal for developer tools, agents, and applications that demand cost-effective, responsive reasoning.
MiniMax-M1 is a large-scale, open-weight reasoning model with 456B total parameters and 45.9B active per token, leveraging a hybrid Mixture-of-Experts (MoE) architecture and a custom "lightning attention" mechanism. It supports context windows up to 1 million tokens and is optimized for long-context understanding, software engineering, agentic tool use, and mathematical reasoning. The model is trained via a custom reinforcement learning pipeline (CISPO) and demonstrates strong performance on FullStackBench, SWE-bench, MATH, GPQA, and TAU-Bench.