MiniMax M2 Overtakes Gemini 2.5 Pro in Open Weight AI Rankings

A Chinese AI challenger moves into the top tier

MiniMax M2 has reached the top of the Artificial Analysis Intelligence Index for open weight models, scoring 61 and placing ahead of Google DeepMind's Gemini 2.5 Pro. As of October 28, 2025, the release puts the Chinese startup among the leading developers of artificial intelligence systems, with a model aimed squarely at coding and software agents that use tools to complete tasks.

MiniMax, a fast rising company backed by major domestic investors, is competing on both capability and cost. Its M2 API costs $0.30 per million input tokens and $1.20 per million output tokens. The company advertises inference throughput near 100 tokens per second and says its pricing is roughly 8 percent of Anthropic's Claude 3.5 Sonnet.

M2's design combines roughly 230 billion total parameters with about 10 billion active at a time. Parameters are the numerical values a model learns during training. By using only part of its network for each token, M2 aims to deliver the capabilities of a much larger system while reducing the computation required to produce answers.

MiniMax has also released the model's full weights under a permissive license. Those weights are the learned values needed to run the model. Developers can download them, adapt the system through fine tuning and deploy it within their own infrastructure, alongside the option of buying access through MiniMax's API.

The release arrives during an increasingly competitive year for Chinese open weight AI. DeepSeek, Alibaba's Qwen, Moonshot AI and other developers have pushed reasoning and long context processing. MiniMax's focus is where businesses are already spending money: software development, research and support workflows that require an AI system to act as well as answer.

What the intelligence ranking measures

Artificial Analysis combines demanding tests of mathematics, science, instruction following, coding and agent workflows into its Intelligence Index. M2's score of 61 places it first among open weight models and among the leading models across the broader field.

Google's Gemini 2.5 Pro falls below M2 in that combined assessment. M2 approaches Anthropic's Claude Sonnet 4.5 on several coding and agent tasks, while the newest proprietary systems, including GPT 5 and xAI's Grok, remain ahead in parts of the comparison. Its position relative to Claude Sonnet 4.5, GPT 5 and Grok 4 varies with the task being measured.

Artificial Analysis also includes knowledge and reasoning assessments such as GPQA Diamond and AIME. Together with the coding tests, these help explain why M2 is attracting attention beyond software development: its strength in programming is accompanied by broader reasoning ability.

Developers weighing the model against proprietary alternatives therefore have several choices to compare. M2 offers performance approaching leading systems in important workflows, low API prices and the freedom to operate the model themselves. Proprietary rivals retain advantages on some tasks, making the intended workload central to any purchasing decision.

Agents that search, act and check their work

MiniMax designed M2 for agents, meaning AI systems that work through a task by planning steps, using tools, reading the results and deciding what to do next. An agent might search the web, query internal information, run code and then check whether its answer or work product is correct.

MiniMax reports scores of 77.2 on tau squared Bench, 44.0 on BrowseComp and 65.5 on FinSearchComp global. These agent assessments cover areas including browsing and citation, financial information retrieval, and planning and recovery when an attempted step fails.

M2's reported strengths include creating a plan, running shell commands, finding and reading information, calling an interpreter to test code, correcting errors and presenting evidence that can be traced to its origin. The intended result is an agent that can keep progressing through a task without requiring a person to direct every action.

MiniMax's approach is especially relevant to research and support work that crosses several systems. A useful answer may depend on information retrieved from a database, a calculation performed in a code runner and a check against another document. M2 is built to connect those actions into a continuing workflow.

The model's performance on BrowseComp and FinSearchComp global is part of its competitive appeal against closed systems. Those tasks put weight on finding information, supplying citations and recovering from mistakes, rather than simply producing a fluent response to a single question.

Coding across a complete development task

MiniMax built M2 for software work that extends beyond generating a small piece of code. Its target workflow includes editing several files, compiling software, running tests and repeatedly repairing problems until the task is complete.

SWE Bench style evaluations and Terminal Bench assess work in software repositories and command line environments. M2 shows competence in these settings across programming languages, supporting its intended role as a coding assistant that can participate in a broader development process.

A command line environment lets software receive instructions through text commands. Within an agent workflow, M2 can use that environment to inspect files, execute programs and read error messages. The next action can then depend on what happened, allowing the model to revise its work instead of stopping after its first attempt.

Businesses considering an internal copilot may also need the same assistant to research a problem or analyse information. M2's results across coding, knowledge and reasoning tests make that combination part of its appeal. Software engineering copilots, data analysis assistants and customer support agents are among the intended uses.

How 230 billion parameters fit a smaller active budget

M2 uses a sparse Mixture of Experts architecture. Instead of passing every token through the entire network, a learned gate selects a few specialised parts, called experts, to process it. The model therefore has a large total capacity but activates a much smaller share for each inference, the process of generating an answer.

MiniMax's technical documentation and model card describe roughly 230 billion total parameters and 10 billion activated per inference. Early descriptions put the total at 200 billion, while the company's technical pages give the larger figure with the same active parameter budget.

The design gives M2 access to the knowledge and specialisation of a large network while bringing the computation used for each token closer to that of a dense model with 10 billion parameters. A dense model uses its full network each time, whereas M2 selects only the relevant experts.

MiniMax and third party testers describe practical serving on as few as four NVIDIA H100 GPUs using FP8 precision. FP8 is a numerical format that uses eight bits to represent values, reducing the memory needed to store them. The deployment recipe puts M2 within reach of midsize engineering teams with access to that hardware.

DeepSeek's recent V3.2 activates about 37 billion parameters per token, while Moonshot AI's Kimi K2 activates around 32 billion. Both figures are substantially higher than M2's 10 billion active parameters. MiniMax's smaller active footprint is intended to reduce computation demands and support faster serving.

MiniMax also targets tail latency and concurrency. Tail latency describes the slowest responses users encounter, while concurrency is the number of requests a system handles at once. Both matter when many people are using an interactive assistant or when several agents are working simultaneously.

Some DeepSeek configurations can require close to 700 gigabytes of memory because of their large total parameter counts. M2's sparse design and serving guidance offer a different operating profile, especially when paired with FP8 and optimised inference frameworks. Teams still need to size memory for the model's precision settings and their intended workloads.

Keeping an agent's working context intact

MiniMax's tool calling guide connects M2 to shells, browsers, retrieval systems and code runners. Tool calling is the mechanism that lets a model request an action from another piece of software, receive its result and continue working.

M2 uses an interleaved thinking format, with visible reasoning traces carried through its interactions. MiniMax encourages preserving those traces across turns so the assistant retains its planning context and evidence chain during longer tasks.

That continuity is intended to help the agent recover from incomplete information or failed steps. It also gives developers material to inspect when investigating why the system took a particular action. Long tasks become easier to debug when their plans, tool results and subsequent decisions remain connected.

MiniMax framed the launch around access to advanced AI. In its public statement, the company described the ambition behind its work:

From the beginning, MiniMax has pursued the vision of 'Intelligence with Everyone.'

MiniMax's team used its public model card to state the model's focus on agents and efficiency:

MiniMax-M2 redefines efficiency for agents.

The company's M2 powered Agent service divides work into two modes. Lightning Mode streams instant responses for simple queries. Pro Mode handles more complex jobs that run longer, including research, data extraction and coding across several stages.

Those modes reflect two different patterns of use: getting a quick answer and handing over a project that needs planning, execution and verification. M2's agent design is aimed at supporting the second pattern without losing the ability to respond quickly to simpler requests.

Access, prices and integration choices

MiniMax offers M2 through downloadable weights and a hosted API. Developers can obtain the model repository on platforms such as Hugging Face, run local inference and fine tune the model for particular domains. Published deployment guides cover SGLang and vLLM, two frameworks used to serve language models.

M2 follows familiar request formats used by OpenAI and Anthropic. That compatibility can reduce the changes needed to connect an existing application to MiniMax's service or to a deployment operated internally.

The API charges $0.30 per million input tokens and $1.20 per million output tokens. Tokens are the units of text processed by a language model, often words or parts of words. Input charges apply to material sent to the system, while output charges apply to the text it generates.

MiniMax's comparison with Claude 3.5 Sonnet puts M2 at roughly 8 percent of that model's price. Its advertised throughput is near 100 tokens per second. Low charges and fast streaming are central to the company's effort to attract developers and enterprises running frequent or extended interactions.

MiniMax has also promoted limited time free access to its Agent application and API. Its documentation provides recommended sampling settings, which control aspects of how responses are generated, and a tool calling schema that specifies how external services connect to the model.

Structured function calling, long context processing and robust tool use are other features relevant to buyers. Structured calls give applications a predictable format for requested actions, while a long context lets a model process more material within an interaction.

Organisations wanting to operate entirely within their own premises can follow the company's serving guidance and choose appropriate GPU memory arrangements for different numerical precision settings. The combination of open weights and familiar API formats gives teams more control over deployment and reduces dependence on a single hosted provider.

Competition turns toward operating costs

MiniMax's ranking adds pressure to competitors in China and the United States. DeepSeek and Qwen have drawn attention through training at the scale of leading models and work on long context processing. M2 is making its case through a smaller active parameter budget, coding ability and agent performance at a lower operating cost.

Enterprises running hundreds of agents, or keeping coding assistants available throughout the working day, must account for repeated requests and lengthy outputs. MiniMax's API pricing undercuts large proprietary rivals by an order of magnitude, making cost a central part of its competitive position.

Lower charges can help a business move from a limited trial to broader use. Open weights provide another option: a team can adapt and run M2 internally instead of relying entirely on a paid external service. For firms handling sensitive information, the ability to inspect, fine tune and deploy the model within their own environment is part of the appeal.

M2's combination of access, cost and capability challenges proprietary platforms that depend on premium pricing. Its strongest case is in workflows where useful reasoning must be paired with dependable tool execution, including software engineering, financial information searches, research and customer support.

What buyers need to test before deployment

Artificial Analysis uses controlled evaluation setups and fixed prompts. A company's own data, software tools and tasks may produce a different pattern of performance. Teams considering M2 should run representative evaluations and measure response times, throughput and agent reliability with the tools they actually intend to use.

MiniMax's recommendation to preserve interleaved thinking traces also creates an operational tradeoff. Longer outputs and retained conversation histories consume more context space. Developers need to manage that space carefully, using trimming or summaries where appropriate while preserving the continuity needed for lengthy tasks.

Teams seeking maximum throughput should test quantisation settings and batching strategies. Quantisation reduces the numerical precision used to represent model values; batching processes several requests together. Both choices affect the balance between speed, cost and accuracy.

Organisations deploying an agent with access to external resources or sensitive information also need monitoring, controls on tool calls and limits on access. M2's open weights allow modification and internal deployment, leaving the operating organisation responsible for its safety and compliance arrangements.

MiniMax's release offers a strong combination for evaluation: a leading open weight ranking, low published API prices and a design built around coding and tools. The practical choice for a buyer depends on how those strengths perform within its own workflows, hardware and security requirements.

At a Glance

• As of October 28, 2025, M2 scores 61 on the Artificial Analysis Intelligence Index and ranks first among open weight models.
• M2 places ahead of Gemini 2.5 Pro in the combined assessment, approaches Claude Sonnet 4.5 on several tasks and trails leading proprietary systems on others.
• MiniMax's documentation gives roughly 230 billion total parameters and 10 billion active per inference, compared with an early total of 200 billion.
• MiniMax reports 77.2 on tau squared Bench, 44.0 on BrowseComp and 65.5 on FinSearchComp global.
• Coding evaluations include SWE Bench style tasks and Terminal Bench, alongside reasoning assessments such as GPQA Diamond and AIME.
• The API costs $0.30 per million input tokens and $1.20 per million output tokens, with advertised throughput near 100 tokens per second.
• Full weights are available under a permissive license, with deployment guidance for SGLang and vLLM and compatibility with common API formats.
• Serving is described on as few as four NVIDIA H100 GPUs at FP8 precision.
• DeepSeek V3.2 activates about 37 billion parameters per token and Kimi K2 about 32 billion, compared with M2's 10 billion.
• Buyers need to test their own workflows and plan for context management, serving performance, tool access controls and safety.