Everyone Missed MiniMax M3 (And I Think That’s About to Change)

MiniMax M3 arrived amid Opus 4.8 hype and GPT 5.6 release rumour, so most people missed out.

M3 is the first open-weight model to combine three capabilities that have only existed in closed-source frontier models: frontier-level coding, a 1M token context window, and native multimodal support covering image, video, and computer use.

On SWE-Bench Pro, M3 scores 59.0%, surpassing both GPT-5.5 and Gemini 3.1 Pro and sitting just below Claude Opus 4.7. On Terminal-Bench 2.1, it hits 66.0%.

On Claw-Eval, an end-to-end framework for autonomous agents, it scores highest among all tested models.

For developers who have been watching open-source models inch closer to frontier performance, M3 is the best so far.

I wanted to break this down for you, so let me walk you through what changed, what the architecture looks like.

M3 Architecture MSA

MiniMax explained why M3 can handle 1M tokens. The answer is MSA, short for MiniMax Sparse Attention.

Standard attention mechanisms have a fundamental problem: as context grows, compute cost grows at a quadratic rate. Double the context, and you quadruple the compute.

This is why most models degrade at long context, and why hitting 1M tokens has been a closed-source privilege.

MSA solves this by replacing full attention with a sparse approach that partitions the key-value cache into blocks more precisely than previous methods like DSA and MoBA.

Each block is read once, memory access stays contiguous, and the arithmetic efficiency improves by more than 4x compared to common open-source sparse attention implementations.

The result is:

• At 1M token context, per-token compute is 1/20th of the previous generation
• Prefilling is more than 9x faster
• Decoding is more than 15x faster

And across multiple ablations, MSA matched full attention on the vast majority of tasks, so you are not trading quality for speed.

This is what makes the 1M context window on M3 different from other models.

Coding and Agentic Capabilities

MiniMax built M3 with the capability of handling real engineering work over extended sessions.

The benchmarks reflect that focus.

• SWE-Bench Pro: 59.0% (ahead of GPT-5.5 and Gemini 3.1 Pro, just below Opus 4.7)
• Terminal-Bench 2.1: 66.0%
• SWE-fficiency: 34.8%
• MCP Atlas: 74.2%

But the two real-world tests MiniMax ran tell a more interesting story than any leaderboard.

1. Paper Reproduction

MiniMax gave M3 an ICLR 2025 Outstanding Paper Award winner and asked it to reproduce the research from scratch.

M3 ran for nearly 12 hours, produced 18 commits and 23 experimental figures, and completed the core experiments without human intervention.

It needed long context to hold the paper, code, and experiment logs together at once, multimodal capability to read the charts and formulas, and strong enough coding skills to carry the task through to the end.

2. CUDA Kernel Optimization

This is important if you work with inference infrastructure.

MiniMax asked M3 to optimize an FP8 matrix multiplication kernel on NVIDIA Hopper GPUs, a task that takes an experienced team one to two weeks. The model started with only a task description and a broken Triton skeleton, with no reference implementation to work from.

Over 24 hours, M3 made 147 benchmark submissions and 1,959 tool calls. It worked through performance plateaus, kept exploring when other models had already stopped, and pushed Hopper FP8 hardware utilization from 7.6% on the first version to 71.3% by the end, a 9.4x speedup.

Most models stopped making progress within the first 30 submissions. M3’s best solution came on submission 145.

The long-horizon persistence is what separates an agentic model from code autocomplete tools, and it is the best demo of MSA’s long-context handling capability.

Multimodality and MiniMax Code

Most models add multimodal support as a feature bolted on after training.

MiniMax trained M3 with mixed-modality data from step zero, meaning text, image, and video were interleaved throughout the entire training process.

Their internal experiments showed that this interleaved approach contributes more to model performance than most teams assumed.

On OmniDocBench, a multimodal benchmark covering document understanding, M3 scores above Gemini 3.1 Pro. On SVG-Bench, which tests visual generation from text and image inputs, it surpasses Claude Opus 4.7.

The practical output of this is computer use.

A user can describe a task on their phone, and MiniMax Code will carry it out on a desktop, moving across applications, files, and systems without supervision.

MiniMax Code

MiniMax Code is the agent product built and trained alongside M3.

It is their equivalent of Claude Code, and it ships with a few specific design choices.

For long tasks, MiniMax Code uses an Agent Team that breaks work into multi-stage, parallel workflows run by a cluster of agents.

A Producer and Verifier loop runs throughout, where agents continuously produce output, review it, and correct course before delivering results. MiniMax says it can run without human intervention for days.

MiniMax Code is built on OpenCode and Pi, both open-source projects, and they plan to open-source MiniMax Code itself in the future.

Pricing and Access

M3 is available through three channels: MiniMax Code, the Token Plan, and the API.

Token Plan

• Plus $20/month: ~1.7B tokens of M3 usage
• Max $50/month: ~5.1B tokens of M3 usage
• Ultra $120/month: ~9.8B tokens of M3 usage

Text, image, speech, and music all draw from the same pool. At those token volumes, MiniMax is positioning this as one of the most generous subscription plans at each price point.

API

API pricing splits by input length.

• Calls under 512K input tokens are billed at the standard rate, which covers most coding and conversation use cases
• Calls above 512K, intended for full-repository analysis and long document parsing, are billed at a higher rate.

Thinking mode can be toggled on or off at request time. On for complex reasoning and agentic tasks, off for faster, latency-sensitive work like code completion. Both modes share the same pricing.

Final Thoughts

M3 is the best open-weight model I have reviewed to date.

The combination of frontier coding, 1M context, and native multimodality in a single open model delivers great value at a good price.

It's important to note that the CUDA kernel test and paper reproduction tasks show M3 can hold a long task together, recover from plateaus, and deliver without someone watching over it.

I tested MiniMax M3 with Claude Code and shared the results here. If you have already run it, let me know what you found in the comments.

Claude Code Masterclass Course

Every day, I’m working hard to build the ultimate Claude Code course, which demonstrates how to create workflows that coordinate multiple agents for complex development tasks. It’s due for release soon.

It will take what you have learned from this article to the next level of complete automation.

New features are added to Claude Code daily, and keeping up is tough.

The course explores Agents, Hooks, advanced workflows, and productivity techniques that many developers may not be aware of.

Once you join, you’ll receive all the updates as new features are rolled out.

This course will cover:

• Advanced subagent patterns and workflows
• Production-ready hook configurations
• MCP server integrations for external tools
• Team collaboration strategies
• Enterprise deployment patterns
• Real-world case studies from my consulting work

If you’re interested in getting notified when the Claude Code course launches, click here to join the early access list →

**(** Currently, I have 70,000+ already signed-up developers)

I’ll share exclusive previews, early access pricing, and bonus materials with people on the list.

Let’s Connect!

If you are new to my content, my name is Joe Njenga

Join thousands of other software engineers, AI engineers, and solopreneurs who read my content daily on Medium and on YouTube where I review the latest AI engineering tools and trends. If you are more curious about my projects and want to receive detailed guides and tutorials, join thousands of other AI enthusiasts in my weekly AI Software engineer newsletter

If you would like to connect directly, you can reach out here:

[embed]AI Automation Software Engineer (10+ Years Experience) My expertise spans LLMs, vector databases, RAG pipelines, MCP servers, and intelligent automation workflows that solve…njengah.com

Follow me on Medium | YouTube Channel | X | LinkedIn | GitHub