MiniMax Open Sources Native Multimodal Flagship Model MiniMax M3

Today we introduce MiniMax-M2.5. After large-scale reinforcement learning training in hundreds of thousands of real complex environments, M2.5 has reached or refreshed industry SOTA in productivity scenarios such as programming, tool calling and search, and office work. Examples include SWE-Bench Verified (80.2%), Multi-SWE-Bench (51.3%), and BrowseComp (76.3%).

M2.5 optimizes the model's ability to decompose complex tasks and token consumption during the thinking process, enabling it to complete complex Agentic tasks faster. For instance, in SWE-Bench Verified testing, it completed tasks 37% faster than the previous version, M2.1.

M2.5 is the first frontier model that can be used infinitely without considering usage costs. At an output speed of 100 tokens per second, M2.5 costs only $1 to work continuously for one hour; at 50 tokens per second, it costs only $0.3. We hope M2.5 can accelerate the development of the Agent era.

Programming: Think and Build Like an Architect

In core programming tests, M2.5 has significantly improved compared to the previous generation, reaching a level similar to the Claude Opus series. On multi-language related tasks like Multi-SWE-Bench, M2.5 achieved first place.

M2.5 possesses the ability to "think and build like an architect." For example, the model has evolved native Spec behavior: before writing code, it actively decomposes functions, structure, and UI design from an architect's perspective, achieving complete preliminary planning.

M2.5 was trained in over 10 languages (including GO, C, C++, TS, Rust, Kotlin, Python, Java, JS, PHP, Lua, Dart, Ruby) and hundreds of thousands of real environments. Not limited to bug fix scenarios, M2.5 performs reliably across the full lifecycle of complex system development: 0-1 system design and environment construction, 1-10 system development, 10-90 feature iteration, and 90-100 comprehensive code review and system testing. It covers full-stack projects across multiple platforms including Web, Android, iOS, Windows, and Mac, including Server-side APIs, functional logic, and databases, rather than just "frontend web demos."

To measure relevant capabilities, we upgraded the VIBE benchmark to a more complex and challenging Pro version: significantly improving task complexity, domain coverage, and evaluation accuracy. Overall, M2.5 performs comparably to Opus4.5.

We focused on the model's generalization across different scaffolds. We tested performance on the SWE-Bench Verified evaluation set across different programming scaffolds. On Droid, M2.5 achieved a pass rate of 79.7, exceeding M2.1's 71.3 and Opus 4.6's 78.9; on OpenCode, M2.5 achieved a pass rate of 76.1, exceeding M2.1's 72.0 and Opus 4.6's 75.9.

Search and Tool Calling

Search and tool calling are prerequisites for models to automatically handle complex tasks. In evaluations on leaderboards such as BrowseComp and Wide Search, M2.5 has reached top industry levels in these areas. Meanwhile, the model's generalization capabilities have also improved. M2.5 demonstrates more stable performance when facing unfamiliar scaffold environments.

In real search tasks by human experts, using search engines is only a small part; more often, it involves deep exploration within professional webpages. For this, we constructed RISE (Realistic Interactive Search Evaluation) to measure the model's search capabilities on real professional tasks. Results show M2.5 performs excellently on real-world expert-level search tasks.

Compared to previous models, M2.5 demonstrates higher decision maturity when handling complex tasks: it has learned to solve problems with more precise search rounds and better token efficiency. For example, in multiple tasks including BrowseComp, Wide Search, and RISE, M2.5 achieved better results with lower round consumption, saving approximately 20% of round consumption compared to M2.1. This indicates the model is no longer just "getting the answer right," but approaching results via a more streamlined path.

Office

We considered how to deliver truly deliverable outputs in office scenarios. For this, we engaged in deep collaboration with senior practitioners in fields such as finance, law, and social sciences. They proposed requirements, provided feedback, participated in standard definition, and directly constructed data, bringing industry tacit knowledge into the model's training process. On this basis, M2.5 achieved significant capability improvements in advanced office scenarios such as Word, PPT, and Excel financial modeling. On the evaluation level, we constructed an internal Cowork Agent evaluation framework (GDPval-MM) to assess delivery quality and trajectory professionalism via pairwise comparison, while monitoring token costs throughout the process to estimate the model's actual benefits in productivity scenarios. In comparisons with mainstream models, it achieved a 59.0% average win rate.

Efficiency

We always hope Agents complete complex tasks in as little time as possible. This depends on the model's ability to decompose complex tasks, the model's token efficiency, and the model's inference speed. Our model originally provides an inference speed of 100TPS, which is nearly double that of current mainstream models. Meanwhile, we focused on optimizing the model's ability to decompose complex tasks and token consumption during the thinking process in reinforcement learning. Combined, these three aspects give M2.5 significant advantages in time and cost for completing complex tasks.

For example, when running the SWE-Bench Verified evaluation set, M2.5 consumed an average of 3.52M tokens per task. In contrast, M2.1 consumed 3.72M tokens. Meanwhile, due to improvements in capabilities such as parallel tool calling, end-to-end runtime decreased from an average of 31.3 minutes to 22.8 minutes, a 37% speed increase. This runtime is basically on par with Claude Opus 4.6's 22.9 minutes.

Cost

The original intention behind designing the M2 series was to enable running complex Agents without cost constraints. As our capabilities continue to improve, we believe M2.5 has nearly achieved this goal. M2.5 offers two versions with the same effect but different speeds and prices: a fast version around 100 TPS, costing only $0.3 per million tokens input and $2.4 per million tokens output. The 50 TPS version has an output price half as low. Based on output pricing, the 50 TPS version is 1/10 to 1/20 the price of models like Opus, Gemini 3 Pro, and GPT5.

At an output speed of 100 tokens per second, continuous work for one hour costs only $1, while at 50 tokens per second, it costs only $0.3. This means $10,000 can allow 4 Agents to work continuously for a year. We believe M2.5 provides the possibility of economically almost unlimited construction and operation of Agents. For the M2 series models, the only remaining question becomes the speed of model capability progress.

Progress Speed

In the past 108 days, we successively updated M2, M2.1, and M2.5, with model progress speed exceeding our original expectations. For example, on SWE-Bench Verified, the most representative benchmark in the programming field, the M2 series models maintained the fastest progress speed in the industry compared to model series like Claude, GPT, and Gemini.

RL Scaling

We believe the core reason for achieving the above progress is large-scale reinforcement learning. It significantly improved model capabilities and generalization to scaffolds and environments. Through co-design of the Agent RL framework, algorithms, Reward design, and engineering optimization, we support efficient optimization for any Agent scaffold and environment. We conducted large-scale training on hundreds of thousands of Agent scaffolds and environments including numerous real internal company tasks, verifying that model capabilities achieve near-linear improvement with scaling of compute and task numbers.

Forge - Native Agent RL Framework

As a native Agent RL framework, Forge completely decouples the underlying training/inference engine from the Agent by introducing an intermediate layer in its design, supporting access for any Agent, allowing us to optimize the model's generalization on Agent scaffolds and tools. To improve system throughput, we optimized asynchronous scheduling strategies to balance system throughput and sample off-policyness, and designed a tree-merged training sample strategy, achieving approximately 40x training acceleration.

Agent RL Algorithms & Reward Design

At the algorithm level, we continued using the CISPO algorithm proposed at the beginning of the year to ensure stability of MoE models during large-scale training. Addressing the credit assignment challenge brought by long contexts in Agent scenarios, we introduced a Process Reward mechanism to monitor completion quality across the entire chain. Additionally, to deeply align with user experience, we directly estimate the time consumption of tasks in real environments as Reward, achieving a better balance between model effectiveness and response speed.

More information on RL scaling and the Agent RL framework Forge will be provided in subsequent tech blogs.

MiniMax Agent: M2.5 as a Real World Coworker

M2.5 has been fully launched in MiniMax Agent, bringing the best Agentic experience.

We distilled the core capability of processing information into standard Office Skills, deeply integrated into the Agent. In All-Rounder (MAX) mode, when handling tasks such as Word layout, PPT editing, and Excel calculations, MiniMax Agent will automatically load corresponding Office Skills based on file type, improving task output quality. Additionally, users can combine Office Skills with industry experience in specific fields to create reusable Experts (Expert) oriented towards specific task scenarios.

Taking industry research as an example, after merging mature research framework SOPs with Word Skills, the Agent can strictly follow the established framework, automatically scrape data, and organize.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.