Zademy

MiniMax M3: One Million Tokens, Native Multimodal Reasoning, and China's Bet on Coding Agents

Tools
MiniMax; M3; LLM; Multimodal; Agents; AI
words words

I was reviewing models for a side project that involves understanding mid-sized repositories and generating structured changes. While comparing options, MiniMax M3 showed up. I wasn't expecting it: the Chinese company had been off my radar for months, except for rumors they were building something big to compete with GPT-4o, Claude 4, and Gemini 2.5 Pro. It was announced in early June and the weights dropped days later, so that same week I started testing it on my own code.

Not because the benchmarks convinced me. Benchmarks exhaust me. I stopped because M3 raises an uncomfortable question: what happens when a model reads your entire codebase, understands the attached documentation, looks at screenshots of the bug, and proposes a patch without you assembling the context piece by piece. That's the promise. And that's what I went to check.

One million tokens, with fine print

I've worked with 128K and 200K token models. They work, but the manual labor stays: copying snippets, summarizing issues, pasting docs, reminding the model where we left off three messages ago. With a million tokens the equation moves. In theory the whole src/ of a mid-sized app fits, plus the API documentation, several issue tracker threads, a couple of bug screenshots, and there's still room left for the answer.

The fine print: that million is the ceiling, not the guarantee. What MiniMax commits to as a minimum is 512K. That's already an enormous context, more than almost anyone else offers, but you want to know the difference before you design a workflow around the round number. And the back-of-envelope conversion (one million tokens is roughly 750,000 English words) is an estimate, not a spec. Treat it as such.

I don't just care that the model understands more. I want to trust that it understands the relationships between all of it. A long context used badly is still noise, just well organized.

My first test was uploading a ~40,000-line TypeScript repository and asking it to identify where shared state between two modules was handled. It got the structure right and missed an edge case that only showed up in a test. That seemed fair to me: it saw the forest. The missing trees I still check myself.

Multimodal from the ground up, but no audio

When a model "sees," the usual setup is text plus a vision module taped on: the model processes words and occasionally receives an image description generated by another system. MiniMax says that in M3 the multimodality is built in from the start: image and video enter the reasoning without passing through that intermediary.

With catalog-level precision, because a lot of enthusiastic summaries are floating around: M3 accepts image and video as input, and responds with text only. It does not accept audio. MiniMax's voice and music live in separate models, and several notes out there mix things up. I misread it myself the first time.

In my tests it understood simple interface screenshots well. With dense screens it started inventing class names that didn't exist, the classic vice of vision models. Even so, the ceiling feels higher than with the separate pipelines I'd used before. For visual debugging it already works: you send the browser screenshot with the error and the model connects it to the part of the code that produces it.

The test on my own code

This is where I came in most skeptical. Making correct changes in an existing system, respecting conventions, tests, and dependencies, remains the biggest open problem in the craft.

First something simple: adding validation to a React form, touching the component, the schema, and a test. The answer was usable. Not perfect: it used a validation library that wasn't in the project and mixed English with Spanish in the error messages. But the structure of the change was right and it saved me time.

Then something more ambitious: moving business logic out of a component into a custom hook. It failed in an interesting way. The hook worked, but it didn't capture a side effect that depended on a DOM event. It respected the shape, not the behavior. That kind of error only gets caught by someone with product context.

To place the moment: MiniMax has published official numbers. 80.5 on SWE-bench Verified, 59.0 on SWE-Bench Pro, 66.0 on Terminal-Bench 2.1, and 83.5 on BrowseComp. I mention them because they exist, not because I take them as truth. They're vendor numbers, measured by whoever is selling the model, and promotional benchmarks always look better than the actual afternoon with the editor open.

My references that June week were Claude 4 Sonnet/Opus for long technical reasoning, GPT-4o/o3 for instruction following and tools, Gemini 2.5 Pro for long context and multimodality (two million tokens), and DeepSeek-V3 for cost-performance. The field has moved since; take it as a June snapshot, not a current table. M3 wants to sit at the intersection of Gemini and Claude: long context, multimodality, and agents. In English and Chinese it did well. In Spanish it was usable, though less refined than Claude or GPT-4o on technical nuance.

Where to get it and what holds me back

You use M3 through the MiniMax API (platform.minimax.io), in MiniMax Code, their desktop agent, with open weights on HuggingFace and GitHub if you'd rather run it yourself, and on OpenRouter too. When I started testing, prices weren't public; they published them later: 0.30and0.30 and 1.20 per million input and output tokens up to 512K of context, a 50% launch discount, and a Token Plan Plus at $20 a month. I leave that as data, not as a recommendation.

What holds me back most isn't the model, it's the ecosystem around it. Claude has Cursor, GPT-4o has Copilot, Gemini has Google's machinery. MiniMax is still building its own outside China. There's also less public information about architecture, training data, and independent evaluations than what Anthropic, OpenAI, or Google publish. And if you work in corporate environments, check where they process your information before sending them a full repository.

For now

MiniMax M3 is a serious proposal. It's not a niche experiment: it aims straight at what we engineers do (understand code, propose changes, reason about systems), and the focus on coding and agents sounds honest, not like marketing that wants to be good at everything.

It isn't a revolution that makes Claude or Gemini obsolete either. It's one more option, strong on long context and multimodality, weak on ecosystem and maturity outside its main market.

I'm still using it on concrete tasks: refactoring, legacy code analysis, test generation, PR review. If something changes my mind, for better or worse, I'll write an update. The recommendation stays simple: try it on your own code before believing any benchmark. Numbers sell, but code doesn't lie.

Resources