MiniMax M3 Officially Released: A 1 Million-Token AI Model Built for Complex Agentic Tasks

Artificial Intelligence (AI) is no longer evolving around the simple question of how well a model can answer a prompt.

The focus is shifting toward something bigger: how much work can AI actually complete on its own?

That shift can be seen in the release of MiniMax M3, the latest AI model from MiniMax designed for complex coding and agentic tasks.

MiniMax M3 combines several capabilities in one model, including a 1 million-token context window, native multimodality, advanced coding capabilities, tool use, and long-horizon autonomous execution.

So, what makes MiniMax M3 worth paying attention to?

MiniMax M3 Is More Than Just an AI Chatbot

MiniMax M3 is designed to handle more complex workflows than simply answering questions.

The model can assist with coding, use tools, process large amounts of information, understand images and videos, and interact with computers to complete tasks.

This brings M3 closer to the concept of an AI agent—an AI system that does not simply provide an answer, but can take multiple actions to accomplish a specific goal.

MiniMax describes M3 as a frontier model built for coding and agentic tasks.

1 Million Tokens: Understanding Extremely Long Contexts

One of the most notable features of MiniMax M3 is its ability to handle up to 1 million tokens of context.

Simply put, a model's context window determines how much information it can process within a single workflow.

With a context window of up to 1 million tokens, M3 can potentially handle tasks involving:

• Large documents
• Entire software repositories or codebases
• Long conversations and instructions
• Experiment logs
• Multiple sources of information

This capability becomes increasingly important when AI is used for tasks that require maintaining context over long periods rather than answering a single, isolated question.

But the interesting part is not just the size of the context window.

MiniMax Introduces a New Attention Architecture

To handle extremely long contexts more efficiently, MiniMax introduced MSA, or MiniMax Sparse Attention.

The architecture is designed to reduce computational requirements when processing very long contexts.

According to MiniMax, at a context length of 1 million tokens, M3's per-token compute is approximately 1/20th of the previous generation.

MiniMax also reports more than 9× faster prefill and more than 15× faster decoding at that context length.

This means the model is not simply capable of processing more information—it is also designed to make long-context processing more efficient.

For developers, this could be especially useful when working with large codebases, lengthy documents, or agent workflows that generate significant amounts of context through tool usage.

M3 Was Built with Native Multimodality

MiniMax M3 is not limited to text.

The model was developed with native multimodality, meaning text, images, and video were incorporated into the training process from the beginning rather than being added as an afterthought.

M3 supports image and video inputs, as well as desktop computer operation.

This allows the model to work with different types of information—not just written instructions or code.

As a result, M3 can potentially support workflows that previously required multiple AI tools working together.

Coding Is One of M3's Main Strengths

Another major focus of MiniMax M3 is software engineering.

MiniMax reports strong performance across several coding and agentic benchmarks, including tasks involving software engineering, terminal execution, and computer use.

This makes M3 interesting not only for developers who want AI to generate code, but also for those looking for an AI assistant that can participate in a broader development workflow.

That can include understanding a codebase, writing code, running tools, debugging, testing solutions, and iterating based on the results.

M3 Was Given a 24-Hour Autonomous Coding Task

One of the most interesting experiments shared by MiniMax involved asking M3 to optimize an FP8 GEMM kernel on NVIDIA Hopper GPUs.

The model was allowed to work for approximately 24 hours without human intervention.

During the process, M3 carried out:

• 147 benchmark submissions
• 1,959 tool calls
• Multiple optimization iterations
• Evaluation of results after each attempt
• Further improvements based on benchmark feedback

The final result was significant.

GPU peak utilization improved from 7.6% to 71.3%, representing approximately a 9.4× performance improvement compared with the starting point.

What makes this experiment particularly interesting is not only the final result, but the process itself.

M3 repeatedly experimented, evaluated what worked and what did not, and changed its approach based on the feedback it received.

This is an important characteristic of an AI agent: it is not simply generating an output—it is iterating on its own work.

M3 Can Also Support Autonomous Research Workflows

MiniMax also tested M3 in a more complex research scenario.

In one experiment, M3 was given an award-winning ICLR 2025 research paper and asked to reproduce its experiments autonomously.

The model worked for nearly 12 hours, produced 18 commits and 23 experimental figures, and successfully reproduced the paper's main experiments.

What makes this interesting is the combination of capabilities required.

The AI had to understand the paper, interpret charts and formulas, work with code, run experiments, evaluate the results, and iterate on its approach.

In other words, the task was no longer simply:

"Write some code."

It was much closer to the workflow of a developer or researcher.

But M3 Is Not the Best at Everything

Despite its strong agentic capabilities, MiniMax M3 still has areas where it does not lead.

One example is autonomous post-training.

On PostTrainBench, a benchmark designed to evaluate a model's ability to autonomously train models from scratch, M3 scored 0.37, below Opus 4.7 at 0.42 and GPT-5.5 at 0.39.

This suggests that while M3 demonstrates strong capabilities for engineering and agentic workflows, it may not necessarily be the best choice for every type of AI research automation.

And this is an important point when evaluating AI models:

A strong benchmark score does not mean a model is the best choice for every task.

What Does MiniMax M3 Mean for Developers and Professionals?

The release of MiniMax M3 highlights an important shift in the development of AI.

AI is gradually moving from generating individual outputs toward executing entire workflows.

For developers, this could mean AI helping to understand codebases, write and test code, use tools, debug problems, and continuously optimize solutions.

For researchers, AI can assist with reading papers, running experiments, analyzing results, and iterating on research workflows.

For non-technical professionals, increasingly capable AI agents could also automate tasks that previously required multiple manual steps.

AI Agents Are Changing the Way We Work

MiniMax M3 shows that the future of AI is not only about "how smart is the model?"

A more important question is becoming:

"How much work can the AI complete on its own?"

Long-context processing, multimodal understanding, tool use, coding, and autonomous iteration are becoming important building blocks for AI systems that can act as digital work partners.

However, humans still play a critical role.

The ability to define goals, provide clear instructions, evaluate AI-generated results, and understand when and how to use AI will become increasingly valuable.

It's Time to Think Beyond AI Chatbots

MiniMax M3 is another example of how AI is evolving from a simple chatbot into an AI agent capable of completing increasingly complex tasks with greater autonomy.

With a context window of up to 1 million tokens, native multimodality, advanced coding capabilities, and long-horizon autonomous execution, M3 represents an interesting development in the current AI landscape.

For students, developers, and professionals, one thing is becoming increasingly clear:

The way we work with AI is changing.

It is no longer just about learning how to get the best answer from AI. It is about learning how to build workflows that allow AI to contribute to a task from beginning to end.

If you want to keep up with the latest developments in AI, technology, and skills relevant to the future of work, follow the latest insights from MAXY Academy.

Explore more insights on AI, technology, and digital skills through MAXY Academy's website and social media channels.

Because in the age of AI, those who keep learning will have more opportunities to keep growing.