MiniMax M2.5 Goes Open Source, Ushering in Era of Affordable AI Agents

MiniMax has launched the M2.5 model, the third iteration in its M2 series, released within just 108 days. The open-source model weights are now accessible on ModelScope, marking significant progress in performance, efficiency, and affordability. It excels in programming, search, and workplace applications. Additionally, it offers a seamless access path—from no-code usage to private deployment—along with a comprehensive guide for tool calling and inference parameter tuning, ushering in an era of low-cost AI agents.

Core Capabilities Deliver Multidimensional Breakthroughs

M2.5 has achieved outstanding results across multiple authoritative benchmarks. It reached 80.2% on SWE-Bench Verified, surpassing GPT-5.2 and nearing Claude Opus4.5. On Multi-SWE-Bench, it scored 51.3%, ranking first in multilingual programming performance. BrowseComp performance hit 76.3%, reflecting strong advantages in search and tool-calling tasks. In programming, the model demonstrates architectural-level planning, covering the full software development lifecycle and enabling cross-platform, full-stack development with better framework generalization than Claude Opus4.6. In search, it reduces the number of interaction rounds by 20%, excelling in expert-level search tasks. For office scenarios, it integrates domain knowledge from finance and law, delivering advanced productivity capabilities and achieving a 59.0% win rate against mainstream models in internal evaluations. At the same time, M2.5 is 37% faster than M2.1, matching Claude Opus4.6 in response time while costing only one-tenth as much.

Rapid Iteration Enabled by Technological Innovation

The swift evolution of M2.5 is driven by three core technological advances: first, the Forge native Agent RL framework, which delivers roughly 40x training acceleration; second, the CISPO algorithm, which ensures stability in large-scale training and resolves the long-context credit assignment problem; and third, an innovative Reward design that balances model performance and response speed. Thanks to these technologies, M2.5 now handles 30% of daily tasks and 80% of new code submissions within MiniMax. Over 108 days, the M2 series’ SWE-Bench Verified score rose from 69.4% to 80.2%, with an iteration pace that leads mainstream industry models.

Multiple Deployment Options to Suit Various Scenarios

M2.5 provides three access methods: no-code, API calls, and local deployment, catering to diverse user needs. Non-technical users can get started immediately using the MiniMax Agent web interface, where over 10,000 reusable “Expert” agents have been created by the community. Developers can utilize the free API on ModelScope or the official API. Two official API versions are available—Lightning and Standard—with costs ranging from 1/10 to 1/20 of comparable models. For local deployment, four solutions are supported: SGLang, vLLM, Transformers, and MLX, each tailored for high-concurrency production, small-to-medium scale deployment, quick validation, or Mac-based local development, complete with hardware requirements and operational instructions.

Dedicated Support for Tool Calls and Parameter Optimization

M2.5 natively supports structured tool calls, enabling parallel invocation of multiple tools. When deployed with vLLM or SGLang, you can directly use the OpenAI SDK format; other frameworks require manual parsing of XML-formatted outputs. A complete workflow and best practices are provided for feeding tool results back into the model. For inference, the recommended parameter settings are temperature=1.0, top_p=0.95, and top_k=40, with flexibility to optimize for specific use cases. Programming prompts can leverage the model’s architectural reasoning, and the model adapts well to over 10 programming languages and various scaffolding tools.