MiniMax M2.5 - MiniMax's next-generation flagship AI model, open sourced from the ground up.

What is M2.5?

M2.5 is MiniMax's lightweight flagship model with 10B activation parameters, emphasizing programming and agentic capabilities. The model supports an ultra-high inference speed of 100 TPS (approximately 3 times that of Claude Opus) and supports full-stack development, complex logic reasoning, and enterprise-level system construction in 10+ languages (Go, Rust, Kotlin, Python, Java, etc.). The model achieves state-of-the-art (SOTA) performance in benchmark tests such as SWE-Bench Verified (80.2%) and Multi-SWE-Bench (51.3%), possessing "architect-level" task decomposition and planning capabilities. The model adopts a native agent architecture and is positioned as the next-generation mainstream model for digital office work.

Main functions of M2.5

• Intelligent ProgrammingIt supports full-stack development in 10+ languages, possesses architect-level planning capabilities, and can complete the entire software lifecycle from system design to test review.
• Agentic execution: Through efficient tool calls and intelligent search, it can autonomously complete complex tasks with lower token consumption and faster speed.
• Office AutomationIt directly delivers professional-grade outputs such as Word, PPT, and Excel, transforming industry tacit knowledge into reusable standardized skills.
• Expert-builtIt supports users in creating custom Experts by integrating domain SOPs and core capabilities. The platform has accumulated 10,000+ experts.
• High cost performanceRunning continuously at 100 TPS for one hour costs only $1, making it economically feasible to run an infinite number of complex agents.

M2.5 Technical Principles

• Forge native Agent RL frameworkBy introducing an intermediate layer to completely decouple the underlying training and push engine from the Agent, it supports the access and optimization of any Agent. Combined with asynchronous scheduling and tree-structured merging training sample strategies, it achieves approximately 40 times training acceleration and verifies that the model's capabilities scale nearly linearly with computing power and the number of tasks.
• Agentic RL Algorithm and Reward DesignThe CISPO algorithm is used to ensure the stability of the MoE model during large-scale training. A process reward mechanism is introduced to solve the problem of credit allocation in long contexts. At the same time, the estimated actual time of the task is used as the reward, achieving an optimal balance between model performance and response speed.
• Efficient Reasoning and Task OptimizationBy efficiently decomposing complex tasks and optimizing token consumption through reinforcement learning incentive models, combined with native 100 TPS inference speed and parallel tool calling capabilities, the completion time of SWE-Bench Verified is reduced by 37% compared to M2.1, while maintaining top-tier performance.
• Large-scale real-world trainingBased on 200,000+ real-world environments, RL training is conducted, covering scenarios such as programming, search, and office work. The real workflows of internal R&D, product, and sales functions are transformed into training data to continuously improve the model's generalization ability in real-world tasks.
• Extreme cost architecture designIt adopts the MoE architecture to achieve high-speed and low-power inference, and provides two versions: 50 TPS and 100 TPS. With the caching mechanism, the output price is reduced to 1/10 to 1/20 of similar models, and for the first time, it realizes the economically feasible unlimited operation of complex agents.

M2.5 performance

• Programming skillsIt achieved a score of 80.2% in the SWE-Bench Verified test, surpassing Claude Opus 4.6; it ranked first in the industry with a score of 51.3% on the Multi-SWE-Bench multi-language programming benchmark; and it performed comparable to Opus 4.5 on the VIBE-Pro internal benchmark. Its cross-scaffolding generalization ability is excellent, achieving a 79.7% pass rate on Droid, exceeding Opus 4.6's 78.9%; and reaching 76.1% on OpenCode, also leading Opus 4.6's 75.9%.
• Search and Tool UsageAchieving a top-tier industry score of 76.3% on the BrowseComp benchmark, Wide Search also leads the pack. It demonstrates expert-level search capabilities in the RISE real-world interactive search evaluation. Compared to its predecessor, M2.1, it saves approximately 20% of search rounds across multiple tasks, significantly improving token efficiency.
• Office SceneIt achieved an average win rate of 59.0% on the internal Cowork Agent evaluation framework GDPval-MM; and achieved excellent results in both the MEWC benchmark based on an Excel competition and the expert-built financial modeling evaluation.

M2.5 project address

• Project official websitehttps://www.minimax.io/models/text

How to use M2.5

• Online experienceVisit the MiniMax Agent website to experience M2.5's conversational and programming capabilities directly.
• API callsDevelopers can register an account and create an API Key at the API call platform https://platform.minimaxi.com to make API calls.