MiniMax-M2.5

<div align="center"> <svg width="60%" height="auto" viewBox="0 0 144 48" fill="none" xmlns="http://www.w3.org/2000/svg"> <path d="M26.6782 7.96523C26.6782 7.02436 25.913 6.26087 24.9739 6.26087C24.0348 6.26087 23.2695 7.0261 23.2695 7.96523V36.2139C23.2695 38.4 21.4904 40.1791 19.3043 40.1791C17.1183 40.1791 15.3391 38.4 15.3391 36.2139V18.0904C15.3391 17.1496 14.5739 16.3861 13.6348 16.3861C12.6956 16.3861 11.9304 17.1513 11.9304 18.0904V25.7722C11.9304 27.9583 10.1513 29.7374 7.96518 29.7374C5.7791 29.7374 4 27.9583 4 25.7722V22.9878C4 22.3635 4.50609 21.8574 5.13043 21.8574C5.75478 21.8574 6.26087 22.3635 6.26087 22.9878V25.7722C6.26087 26.713 7.02605 27.4765 7.96518 27.4765C8.90431 27.4765 9.66954 26.7113 9.66954 25.7722V18.0904C9.66954 15.9044 11.4487 14.1252 13.6348 14.1252C15.8209 14.1252 17.6 15.9044 17.6 18.0904V36.2139C17.6 37.1548 18.3652 37.9183 19.3043 37.9183C20.2435 37.9183 21.0087 37.153 21.0087 36.2139V25.1322V7.96523C21.0087 5.77914 22.7878 4 24.9739 4C27.16 4 28.9391 5.77914 28.9391 7.96523V31.3565C28.9391 31.9809 28.433 32.487 27.8087 32.487C27.1843 32.487 26.6782 31.9809 26.6782 31.3565V7.96523ZM47.6539 14.1252C45.4678 14.1252 43.6887 15.9044 43.6887 18.0904V33.2296C43.6887 34.1704 42.9235 34.9339 41.9843 34.9339C41.0452 34.9339 40.28 34.1687 40.28 33.2296V7.96523C40.28 5.77914 38.5008 4 36.3148 4C34.1287 4 32.3496 5.77914 32.3496 7.96523V40.0348C32.3496 40.9756 31.5843 41.7391 30.6452 41.7391C29.7061 41.7391 28.9409 40.9739 28.9409 40.0348V36.0643C28.9409 35.44 28.4348 34.9339 27.8104 34.9339C27.1861 34.9339 26.68 35.44 26.68 36.0643V40.0348C26.68 42.2209 28.4591 44 30.6452 44C32.8313 44 34.6104 42.2209 34.6104 40.0348V7.96523C34.6104 7.02436 35.3756 6.26087 36.3148 6.26087C37.2539 6.26087 38.0191 7.0261 38.0191 7.96523V33.2296C38.0191 35.4156 39.7982 37.1948 41.9843 37.1948C44.1704 37.1948 45.9496 35.4156 45.9496 33.2296V18.0904C45.9496 17.1496 46.7148 16.3861 47.6539 16.3861C48.593 16.3861 49.3582 17.1513 49.3582 18.0904V31.3565C49.3582 31.9809 49.8643 32.487 50.4887 32.487C51.113 32.487 51.6191 31.9809 51.6191 31.3565V18.0904C51.6191 15.9044 49.84 14.1252 47.6539 14.1252Z" fill="url(#paint0_linear_17_483)"/> <path d="M68.7671 16.5615H71.2541C71.3254 16.5615 71.3845 16.5859 71.435 16.6363C71.4836 16.6868 71.5097 16.7459 71.5097 16.8172V31.1824C71.5097 31.2537 71.4854 31.3128 71.435 31.3633C71.3845 31.4137 71.3254 31.4381 71.2541 31.4381H68.7671C68.6958 31.4381 68.6367 31.4137 68.5862 31.3633C68.5358 31.3146 68.5115 31.2537 68.5115 31.1824V21.812C68.5115 21.7563 68.4976 21.7268 68.4697 21.7268C68.4419 21.7268 68.4123 21.7476 68.3845 21.7911L66.1323 25.318C66.061 25.4311 65.9619 25.4885 65.8349 25.4885H64.581C64.4541 25.4885 64.3549 25.4328 64.2836 25.318L62.0315 21.7911C62.0036 21.7494 61.9741 21.7302 61.9462 21.7372C61.9184 21.7441 61.9045 21.7772 61.9045 21.8328V31.1824C61.9045 31.2537 61.8802 31.3128 61.8297 31.3633C61.7793 31.4137 61.7202 31.4381 61.6489 31.4381H59.1619C59.0906 31.4381 59.0315 31.4137 58.981 31.3633C58.9306 31.3146 58.9062 31.2537 58.9062 31.1824V16.8172C58.9062 16.7459 58.9306 16.6868 58.981 16.6363C59.0315 16.5859 59.0906 16.5615 59.1619 16.5615H61.6489C61.7758 16.5615 61.8749 16.6189 61.9462 16.732L65.1341 21.6833C65.1758 21.7685 65.2193 21.7685 65.261 21.6833L68.4697 16.732C68.541 16.6189 68.6402 16.5615 68.7671 16.5615Z" fill="currentColor"/> <path d="M74.1764 31.3633C74.1259 31.3146 74.1016 31.2537 74.1016 31.1824V16.8172C74.1016 16.7459 74.1259 16.6868 74.1764 16.6363C74.2268 16.5859 74.2859 16.5615 74.3572 16.5615H76.8442C76.9155 16.5615 76.9746 16.5859 77.0251 16.6363C77.0737 16.6868 77.0998 16.7459 77.0998 16.8172V31.1824C77.0998 31.2537 77.0755 31.3128 77.0251 31.3633C76.9746 31.4137 76.9155 31.4381 76.8442 31.4381H74.3572C74.2859 31.4381 74.2268 31.4137 74.1764 31.3633Z" fill="currentColor"/> <path d="M88.3066 16.6361C88.3553 16.5874 88.4162 16.5613 88.4875 16.5613H90.9744C91.0457 16.5613 91.1049 16.5857 91.1553 16.6361C91.204 16.6865 91.2301 16.7457 91.2301 16.817V31.1822C91.2301 31.2535 91.2057 31.3126 91.1553 31.363C91.1049 31.4135 91.0457 31.4378 90.9744 31.4378H88.5727C88.4301 31.4378 88.331 31.3822 88.2753 31.2674L82.771 22.1717C82.7431 22.13 82.7136 22.1109 82.6858 22.1178C82.6579 22.1248 82.644 22.1578 82.644 22.2135L82.6858 31.1805C82.6858 31.2518 82.6614 31.3109 82.611 31.3613C82.5606 31.4117 82.5014 31.4361 82.4301 31.4361H79.9431C79.8718 31.4361 79.8127 31.4117 79.7623 31.3613C79.7118 31.3126 79.6875 31.2518 79.6875 31.1805V16.8152C79.6875 16.7439 79.7118 16.6848 79.7623 16.6344C79.8127 16.5839 79.8718 16.5596 79.9431 16.5596H82.3449C82.4858 16.5596 82.5849 16.617 82.6423 16.73L88.124 25.7822C88.1518 25.8239 88.1797 25.8431 88.2092 25.8361C88.2371 25.8292 88.251 25.7978 88.251 25.7404L88.2301 16.8152C88.2301 16.7439 88.2545 16.6848 88.3049 16.6344L88.3066 16.6361Z" fill="currentColor"/> <path d="M93.8951 31.3633C93.8446 31.3146 93.8203 31.2537 93.8203 31.1824V16.8172C93.8203 16.7459 93.8446 16.6868 93.8951 16.6363C93.9455 16.5859 94.0047 16.5615 94.076 16.5615H96.5629C96.6342 16.5615 96.6934 16.5859 96.7438 16.6363C96.7925 16.6868 96.8186 16.7459 96.8186 16.8172V31.1824C96.8186 31.2537 96.7942 31.3128 96.7438 31.3633C96.6934 31.4137 96.6342 31.4381 96.5629 31.4381H94.076C94.0047 31.4381 93.9455 31.4137 93.8951 31.3633Z" fill="currentColor"/> <path d="M109.267 16.5615H111.754C111.825 16.5615 111.885 16.5859 111.935 16.6363C111.984 16.6868 112.01 16.7459 112.01 16.8172V31.1824C112.01 31.2537 111.985 31.3128 111.935 31.3633C111.885 31.4137 111.825 31.4381 111.754 31.4381H109.267C109.196 31.4381 109.137 31.4137 109.086 31.3633C109.036 31.3146 109.011 31.2537 109.011 31.1824V21.812C109.011 21.7563 108.998 21.7268 108.97 21.7268C108.942 21.7268 108.912 21.7476 108.885 21.7911L106.632 25.318C106.561 25.4311 106.462 25.4885 106.335 25.4885H105.081C104.954 25.4885 104.855 25.4328 104.784 25.318L102.531 21.7911C102.504 21.7494 102.474 21.7302 102.446 21.7372C102.418 21.7441 102.405 21.7772 102.405 21.8328V31.1824C102.405 31.2537 102.38 31.3128 102.33 31.3633C102.279 31.4137 102.22 31.4381 102.149 31.4381H99.6619C99.5906 31.4381 99.5315 31.4137 99.481 31.3633C99.4306 31.3146 99.4062 31.2537 99.4062 31.1824V16.8172C99.4062 16.7459 99.4306 16.6868 99.481 16.6363C99.5315 16.5859 99.5906 16.5615 99.6619 16.5615H102.149C102.276 16.5615 102.375 16.6189 102.446 16.732L105.634 21.6833C105.676 21.7685 105.719 21.7685 105.761 21.6833L108.97 16.732C109.041 16.6189 109.14 16.5615 109.267 16.5615Z" fill="currentColor"/> <path d="M123.782 31.2241L123.144 29.1424C123.116 29.0867 123.079 29.0572 123.038 29.0572H117.81C117.768 29.0572 117.732 29.085 117.704 29.1424L117.088 31.2241C117.046 31.3668 116.954 31.4363 116.812 31.4363H114.112C114.027 31.4363 113.963 31.412 113.921 31.3615C113.879 31.3128 113.871 31.2381 113.9 31.1389L118.49 16.7737C118.532 16.6328 118.624 16.5615 118.766 16.5615H122.102C122.243 16.5615 122.335 16.6328 122.379 16.7737L126.968 31.1389C126.982 31.1668 126.989 31.2033 126.989 31.245C126.989 31.372 126.911 31.4363 126.756 31.4363H124.057C123.916 31.4363 123.824 31.365 123.78 31.2241H123.782ZM118.554 26.7407H122.295C122.38 26.7407 122.408 26.6989 122.38 26.6137L120.467 20.3024C120.453 20.2467 120.432 20.2207 120.403 20.2276C120.375 20.2346 120.352 20.2589 120.339 20.3024L118.469 26.6137C118.455 26.6989 118.483 26.7407 118.554 26.7407Z" fill="currentColor"/> <path d="M128.222 31.353C128.18 31.2974 128.187 31.2261 128.243 31.1409L132.365 24.0643C132.393 24.0226 132.393 23.9791 132.365 23.9374L128.243 16.8609L128.201 16.7339C128.201 16.6209 128.28 16.5635 128.434 16.5635H131.133C131.274 16.5635 131.38 16.6209 131.452 16.7339L134.213 21.6C134.255 21.6852 134.299 21.6852 134.34 21.6L137.102 16.7339C137.173 16.6209 137.28 16.5635 137.42 16.5635H140.099C140.198 16.5635 140.269 16.5913 140.311 16.6487C140.353 16.7061 140.346 16.7756 140.29 16.8609L136.168 23.9374C136.154 23.9791 136.154 24.0226 136.168 24.0643L140.29 31.1409L140.332 31.2678C140.332 31.3809 140.253 31.4383 140.099 31.4383H137.42C137.278 31.4383 137.172 31.3826 137.102 31.2678L134.34 26.4226C134.299 26.3374 134.255 26.3374 134.213 26.4226L131.429 31.2678C131.358 31.3809 131.252 31.4383 131.111 31.4383H128.433C128.333 31.4383 128.262 31.4104 128.22 31.353H128.222Z" fill="currentColor"/> <defs> <linearGradient id="paint0_linear_17_483" x1="3.99826" y1="24" x2="51.6208" y2="24" gradientUnits="userSpaceOnUse"> <stop stop-color="#E21680"/> <stop offset="1" stop-color="#FF633A"/> </linearGradient> </defs> </svg> </div> <hr> <div align="center" style="line-height: 1.4; font-size:16px; margin-top: 30px;"> Join Our <a href="https://platform.minimaxi.com/docs/faq/contact-us" target="_blank" style="font-size:17px; margin: 2px;"> 💬 WeChat </a> | <a href="https://discord.com/invite/hvvt8hAye6" target="_blank" style="font-size:17px; margin: 2px;"> 🧩 Discord </a> community. </div> <div align="center" style="line-height: 1.2; font-size:16px;"> <a href="https://agent.minimax.io/" target="_blank" style="display: inline-block; margin: 4px;"> MiniMax Agent </a> | <a href="https://platform.minimax.io/docs/guides/text-generation" target="_blank" style="display: inline-block; margin: 4px;"> ⚡️ API </a> | <a href="https://github.com/MiniMax-AI/MiniMax-MCP" style="display: inline-block; margin: 4px;"> MCP </a> | <a href="https://www.minimax.io" target="_blank" style="display: inline-block; margin: 4px;"> MiniMax Website </a> </div> <div align="center" style="line-height: 1.2; font-size:16px; margin-bottom: 30px;"> <a href="https://huggingface.co/MiniMaxAI" target="_blank" style="margin: 2px;"> 🤗 Hugging Face </a> | <a href="https://github.com/MiniMax-AI/MiniMax-M2.1" target="_blank" style="margin: 2px;"> 🐙 GitHub </a> | <a href="https://www.modelscope.cn/organization/MiniMax" target="_blank" style="margin: 2px;"> 🤖️ ModelScope </a> | <a href="https://github.com/MiniMax-AI/MiniMax-M2.5/blob/main/LICENSE" style="margin: 2px;"> 📄 License: Modified-MIT </a> </div> <p align="center"> <picture> <img class="hidden dark:block" width="100%" src="figures/bench_11.png"> <img class="dark:hidden" width="100%" src="figures/bench_12.png"> </picture> </p>

Today we're introducing our latest model, MiniMax-M2.5.

Extensively trained with reinforcement learning in hundreds of thousands of complex real-world environments, M2.5 is SOTA in coding, agentic tool use and search, office work, and a range of other economically valuable tasks, boasting scores of 80.2% in SWE-Bench Verified, 51.3% in Multi-SWE-Bench, and 76.3% in BrowseComp (with context management).

Trained to reason efficiently and decompose tasks optimally, M2.5 exhibits tremendous speed in performing complicated agentic tasks, completing the SWE-Bench Verified evaluation 37% faster than M2.1, matching the speed of Claude Opus 4.6.

M2.5 is the first frontier model where users do not need to worry about cost, delivering on the promise of intelligence too cheap to meter. It costs just $1 to run the model continuously for an hour at a rate of 100 tokens per second. At 50 tokens per second, the cost drops to $0.30. We hope that the speed and cost effectiveness of M2.5 enable innovative new agentic applications.

Coding

In programming evaluations, MiniMax-M2.5 saw substantial improvements compared to previous generations, reaching SOTA levels. The performance of M2.5 in multilingual tasks is especially pronounced.

<p align="center"> <picture> <img class="hidden dark:block" width="100%" src="figures/bench_2.png"> <img class="dark:hidden" width="100%" src="figures/bench_1.png"> </picture> </p>

A significant improvement from previous generations is M2.5's ability to think and plan like an architect. The Spec-writing tendency of the model emerged during training: before writing any code, M2.5 actively decomposes and plans the features, structure, and UI design of the project from the perspective of an experienced software architect.

M2.5 was trained on over 10 languages (including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby) across more than 200,000 real-world environments. Going far beyond bug-fixing, M2.5 delivers reliable performance across the entire development lifecycle of complex systems: from 0-to-1 system design and environment setup, to 1-to-10 system development, to 10-to-90 feature iteration, and finally 90-to-100 comprehensive code review and system testing. It covers full-stack projects spanning multiple platforms including Web, Android, iOS, and Windows, encompassing server-side APIs, business logic, databases, and more, not just frontend webpage demos.

To evaluate these capabilities, we also upgraded the VIBE benchmark to a more complex and challenging Pro version, significantly increasing task complexity, domain coverage, and evaluation accuracy. Overall, M2.5 performs on par with Opus 4.5.

<p align="center"> <picture> <img class="hidden dark:block" width="100%" src="figures/bench_4.png"> <img class="dark:hidden" width="100%" src="figures/bench_3.png"> </picture> </p>

We focused on the model's ability to generalize across out-of-distribution harnesses. We tested performance on the SWE-Bench Verified evaluation set using different coding agent harnesses.

• On Droid: 79.7(M2.5) > 78.9(Opus 4.6)
• On OpenCode: 76.1(M2.5) > 75.9(Opus 4.6)

Search and Tool calling

<p align="center"> <picture> <img class="hidden dark:block" width="100%" src="figures/bench_6.png"> <img class="dark:hidden" width="100%" src="figures/bench_5.png"> </picture> </p>

Effective tool calling and search are prerequisites for a model's ability to autonomously handle more complex tasks. In evaluations on benchmarks such as BrowseComp and Wide Search, M2.5 achieved industry-leading performance. At the same time, the model's generalization has also improved — M2.5 demonstrates more stable performance when facing unfamiliar scaffolding environments.

In research tasks performed by professional human experts, using a search engine is only a small part of the process; most of the work involves deep exploration across information-dense webpages. To address this, we built RISE (Realistic Interactive Search Evaluation) to measure a model's search capabilities on real-world professional tasks. The results show that M2.5 excels at expert-level search tasks in real-world settings.

Compared to its predecessors, M2.5 also demonstrates much better decision-making when handling agentic tasks: it has learned to solve problems with more precise search rounds and better token efficiency. For example, across multiple agentic tasks including BrowseComp, Wide Search, and RISE, M2.5 achieved better results with fewer rounds, using approximately 20% fewer rounds compared to M2.1. This indicates that the model is no longer just getting the answer right, but is also reasoning towards results in more efficient paths.

Office work

M2.5 was trained to produce truly deliverable outputs in office scenarios. To this end, we engaged in thorough collaboration with senior professionals in fields such as finance, law, and social sciences. They designed requirements, provided feedback, participated in defining standards, and directly contributed to data construction, bringing the tacit knowledge of their industries into the model's training pipeline. Based on this foundation, M2.5 has achieved significant capability improvements in high-value workspace scenarios such as Word, PowerPoint, and Excel financial modeling. On the evaluation side, we built an internal Cowork Agent evaluation framework (GDPval-MM) that assesses both the quality of the deliverable and the professionalism of the agent's trajectory through pairwise comparisons, while also monitoring token costs across the entire workflow to estimate the model's real-world productivity gains. In comparisons against other mainstream models, it achieved an average win rate of 59.0%.

<p align="center"> <picture> <img class="hidden dark:block" width="100%" src="figures/bench_8.png"> <img class="dark:hidden" width="100%" src="figures/bench_7.png"> </picture> </p>

Efficiency

Because the real world is full of deadlines and time constraints, task completion speed is a practical necessity. The time it takes a model to complete a task depends on its task decomposition effectiveness, token efficiency, and inference speed. M2.5 is served natively at a rate of 100 tokens per second, which is nearly twice that of other frontier models. Further, our reinforcement learning setup incentivizes the model to reason efficiently and break down tasks optimally. Due to these three factors, M2.5 delivers a significant time savings in complex task completion.

For example, when running SWE-Bench Verified, M2.5 consumed an average of 3.52 million tokens per task. In comparison, M2.1 consumed 3.72M tokens. Meanwhile, thanks to improvements in capabilities such as parallel tool calling, the end-to-end runtime decreased from an average of 31.3 minutes to 22.8 minutes, representing a 37% speed improvement. This runtime is on par with Claude Opus 4.6's 22.9 minutes, while the total cost per task is only 10% that of Claude Opus 4.6.

Cost

Our goal in designing the M2-series of foundation models is to power complex agents without having to worry about cost. We believe that M2.5 is close to realizing this goal. We’re releasing two versions of the model, M2.5 and M2.5-Lightning, that are identical in capability but differ in speed. M2.5-Lightning has a steady throughput of 100 tokens per second, which is two times faster than other frontier models, and costs $0.3 per million input tokens and $2.4 per million output tokens. M2.5, which has a throughput of 50 tokens per second, costs half that. Both model versions support caching. Based on output price, the cost of M2.5 is one-tenth to one-twentieth that of Opus, Gemini 3 Pro, and GPT-5.

At a rate of 100 output tokens per second, running M2.5 continuously for an hour costs $1. At a rate of 50 TPS, the price drops to $0.3. To put that into perspective, you can have four M2.5 instances running continuously for an entire year for $10,000. We believe that M2.5 provides virtually limitless possibilities for the development and operation of agents in the economy. For the M2-series, the only problem that remains is how to continually push the frontier of model capability.

Improvement Rate

Over the three and a half months from late October to now, we have successively released M2, M2.1, and M2.5, with the pace of model improvement exceeding our original expectations. For instance, in the highly-regarded SWE-Bench Verified benchmark, the rate of progress of the M2-series has been significantly faster than that of peers such as the Claude, GPT, and Gemini model families.

<p align="center"> <img width="100%" src="figures/bench_10.png"> </p>

RL Scaling

One of the key drivers of the aforementioned developments is the scaling of reinforcement learning. As we train our models, we also benefit from their abilities. Most of the tasks and workspaces that we perform in our company have been made into training environments for RL. To date, there are already hundreds of thousands of such environments. At the same time, we did plenty of work on our agentic RL framework, algorithms, reward signals, and infrastructure engineering to support the continued scaling of our RL training.

Forge –– Agent-Native RL Framework

We designed an agent-native RL framework in-house, called Forge, which introduces an intermediary layer that fully decouples the underlying training-inference engine from the agent, supporting the integration of arbitrary agents and enabling us to optimize the model's generalization across agent scaffolds and tools. To improve system throughput, we optimized asynchronous scheduling strategies to balance system throughput against sample off-policyness, and designed a tree-structured merging strategy for training samples, achieving approximately 40x training speedup.

<p align="center"> <img width="60%" src="figures/rl_1.png"> </p>

Agentic RL Algorithm and Reward Design

On the algorithm side, we continued using the CISPO algorithm we proposed at the beginning of last year to ensure the stability of MoE models during large-scale training. To address the credit assignment challenge posed by long contexts in agent rollouts, we introduced a process reward mechanism for end-to-end monitoring of generation quality. Furthermore, to deeply align with user experience, we evaluated task completion time through agent trajectories, achieving an optimal trade-off between model intelligence and response speed.

<p align="center"> <img width="60%" src="figures/rl_2.png"> </p>

We will release a more comprehensive introduction to RL scaling soon in a separate technical blogpost.

MiniMax Agent: M2.5 as a Professional Employee

M2.5 has been fully deployed in MiniMax Agent, delivering the best agentic experience.

We have distilled core information-processing capabilities into standardized Office Skills deeply integrated within MiniMax Agent. In MAX mode, when handling tasks such as Word formatting, PowerPoint editing, and Excel calculations, MiniMax Agent automatically loads the corresponding Office Skills based on file type, improving the quality of task outputs.

Furthermore, users can combine Office Skills with domain-specific industry expertise to create reusable Experts tailored to specific task scenarios.

Take industry research as an example: by merging a mature research framework SOP (standard operating procedure) with Word Skills, the Agent can strictly follow the established framework to automatically fetch data, organize analytical logic, and output properly formatted research reports — rather than merely generating a raw block of text. In financial modeling scenarios, by combining an organization's proprietary modeling standards with Excel Skills, the Agent can follow specific risk control logic and calculation standards to automatically generate and validate complex financial models, rather than simply outputting a basic spreadsheet.

To date, users have built over 10,000 Experts on MiniMax Agent, and this number is still growing rapidly. MiniMax has also built multiple sets of deeply optimized, ready-to-use Expert suites on MiniMax Agent for high-frequency scenarios such as office work, finance, and programming.

MiniMax itself has been among the first to benefit from M2.5's capabilities. Throughout the company's daily operations, 30% of overall tasks are autonomously completed by M2.5, spanning functions including R&D, product, sales, HR, and finance — and the penetration rate continues to rise. Performance in coding scenarios has been particularly notable, with M2.5-generated code accounting for 80% of newly committed code.

How to Use

MiniMax Agent: https://agent.minimax.io/

MiniMax API Platform: https://platform.minimax.io/

MiniMax Coding Plan: https://platform.minimax.io/subscribe/coding-plan

Local Deployment Guide

Download the model from HuggingFace repository: https://huggingface.co/MiniMaxAI/MiniMax-M2.5

We recommend using the following inference frameworks (listed alphabetically) to serve the model:

SGLang

We recommend using SGLang to serve MiniMax-M2.5. Please refer to our SGLang Deployment Guide.

vLLM

We recommend using vLLM to serve MiniMax-M2.5. Please refer to our vLLM Deployment Guide.

Transformers

We recommend using Transformers to serve MiniMax-M2.5. Please refer to our Transformers Deployment Guide.

KTransformers

We recommend using KTransformers to serve MiniMax-M2.5. Please refer to KTransformers Deployment Guide

ModelScope

You also can get model weights from modelscope.

Inference Parameters

We recommend using the following parameters for best performance: temperature=1.0, top_p = 0.95, top_k = 40. Default system prompt:

You are a helpful assistant. Your name is MiniMax-M2.5 and is built by MiniMax.

Tool Calling Guide

Please refer to our Tool Calling Guide.

Contact Us

Contact us at [email protected].

Appendix

Further benchmark results of M2.5:

Evaluation methods:

• SWE benchmark: SWE-bench Verified, SWE-bench Multilingual, SWE-bench-pro, and Multi-SWE-bench were tested on internal infrastructure using Claude Code as the scaffolding, with the default system prompt overridden, and results averaged over 4 runs. Additionally, SWE-bench Verified was also evaluated on the Droid and Opencode scaffoldings using the default prompt.
• Terminal Bench 2: We tested Terminal Bench 2 using Claude Code 2.0.64 as the evaluation scaffolding. We modified the Dockerfiles of some problems to ensure the correctness of the problems themselves, uniformly expanded sandbox specifications to 8-core CPU and 16 GB memory, set the timeout uniformly to 7,200 seconds, and equipped each problem with a basic toolset (ps, curl, git, etc.). While not retrying on timeouts, we added a detection mechanism for empty scaffolding responses, retrying tasks whose final response was empty to handle various abnormal interruption scenarios. Final results are averaged over 4 runs.
• VIBE-Pro: Internal benchmark. Uses Claude Code as the scaffolding to automatically verify the interaction logic and visual effects of programs. All scores are computed through a unified pipeline that includes a requirements set, containerized deployment, and a dynamic interaction environment. Final results are averaged over 3 runs.
• BrowseComp: Uses the same agent framework as WebExplorer (Liu et al., 2025). When token usage exceeds 30% of the maximum context, all history is discarded.
• Wide Search: Uses the same agent framework as WebExplorer (Liu et al., 2025).
• RISE: Internal benchmark. Contains real questions from human experts, evaluating the model's multi-step information retrieval and reasoning capabilities when combined with complex web interactions. A Playwright-based browser tool suite is added on top of the WebExplorer (Liu et al., 2025) agent framework.
• GDPval-MM: Internal benchmark. Based on the open-source GDPval test set, using a custom agentic evaluation framework where an LLM-as-a-judge performs pairwise win/tie/loss judgments on complete trajectories. Average token cost per task is calculated based on each vendor's official API pricing (without caching).
• MEWC: Internal benchmark. Built on MEWC (Microsoft Excel World Championship), comprising 179 problems from the main and other regional divisions of Excel esports competitions from 2021–2026. It evaluates the model's ability to understand competition Excel spreadsheets and use Excel tools to complete problems. Scores are calculated by comparing output and answer cell values one by one.
• Finance Modeling: Internal benchmark. Primarily contains financial modeling problems constructed by industry experts, involving end-to-end research and analysis tasks performed via Excel tools. Each problem is scored using expert-designed rubrics. Final results are averaged over 3 runs.
• AIME25 ~ AA-LCR: Obtained through internal testing based on the public evaluation sets and evaluation methods covered by the Artificial Analysis Intelligence Index leaderboard.