MiniMax released the M2.5 model, the third version in its M2 series within 108 days. The model's open-source weights are now available on ModelScope, achieving breakthroughs in capability, efficiency, and cost. It performs well in programming, search, and office scenarios. It also offers a full-process access solution from no-code usage to private deployment, as well as a complete practical guide for tool calls and inference parameter optimization, driving the era of low-cost Agents.

image.png

Core Capabilities Achieve Multidimensional Breakthroughs

M2.5 has shown impressive results in multiple authoritative evaluations. SWE-Bench Verified reached 80.2%, surpassing GPT-5.2 and approaching Claude Opus4.5. Multi-SWE-Bench achieved 51.3%, ranking first in multilingual programming capabilities. BrowseComp reached 76.3%, with significant advantages in search and tool calling capabilities. In programming, the model demonstrates architectural-level planning capabilities, covering the entire software development lifecycle, supporting full-stack development across multiple platforms, with better framework generalization than Claude Opus4.6; in search, it reduces 20% of the rounds, performing excellently in expert-level search tasks; in office scenarios, it integrates industry knowledge such as finance and law, showing advanced office capabilities, with an internal evaluation win rate of 59.0% against mainstream models. At the same time, M2.5 is 37% faster than M2.1, matching the time of Claude Opus4.6, with a cost only 1/10 of that.

Technological Innovation Enables Rapid Iteration

The rapid evolution of M2.5 stems from three core technological innovations: first, the Forge native Agent RL framework, which achieves about 40 times training acceleration; second, the CISPO algorithm ensures the stability of large-scale training, solving the long context credit allocation problem; third, the innovative Reward design balances model performance and response speed. This technology enables 30% of daily tasks and 80% of new code submissions within MiniMax to be completed by M2.5. Within 108 days, the SWE-Bench Verified score of the M2 series jumped from 69.4% to 80.2%, with iteration speed leading industry mainstream models.

Multiple Deployment Methods Adapt to Different Scenarios

M2.5 provides three access methods: no-code, API calls, and local deployment, meeting different user needs. Non-technical users can use the MiniMax Agent web interface out-of-the-box, with over 10,000 reusable "Experts" created by users on the platform. Developers can call the free API or official API on ModelScope. The official also launched two API versions: Lightning and Standard, with costs only 1/10 to 1/20 of similar models. Local deployment supports four solutions: SGLang, vLLM, Transformers, and MLX, each suitable for high-concurrency production, small and medium-scale production, quick verification, and Mac local development, and provides hardware requirements and operation steps for each solution.

Tool Calls and Parameter Optimization Have Dedicated Solutions

M2.5 natively supports structured tool calls, allowing parallel calls to multiple tools. With vLLM/SGLang deployment, you can directly use the OpenAI SDK format; other frameworks require manual parsing of XML format output. It also provides a complete process and best practices for returning tool results to the model. In terms of inference, the official recommends a parameter configuration of temperature=1.0, top_p=0.95, top_k=40, and different scenarios can flexibly optimize parameters. Programming prompts can leverage the model's architect thinking, and the model has excellent adaptability to more than 10 programming languages and various scaffolding tools.