MiniMax M3: a powerful new AI you might soon run yourself — explained

China's AI labs keep shipping capable models for a fraction of the usual price, and MiniMax M3 is the latest. Here's what it is, what makes it interesting, and — importantly — the two things you should know before getting too excited about the headlines.

Three things stand out about M3:

1. It's strong at coding. MiniMax says it scores 59.0% on a benchmark called SWE-Bench Pro, which tests whether an AI can write real fixes for real software bugs. That number would put it ahead of rivals like GPT-5.5 and Gemini 3.1 Pro on that specific test — though as we'll cover below, it's MiniMax's own figure, not yet checked by outsiders.

2. It has a very large 'memory'. M3 can take in about one million tokens at once — that's roughly 750,000 words, or an entire codebase or large document collection. Most AI models you use every day handle far less. This matters for anyone using AI to work through lengthy contracts, research papers or large programming projects.

3. It understands images and video, not just text. Rather than being text-only, M3 was built from the start to handle images and video alongside words. This is increasingly standard for top models, but it's worth flagging as a baseline capability.

The price is the part that's unambiguously real: M3 launched on OpenRouter at roughly $0.30 per million tokens for input and $1.20 for output — a promotional rate, but one that's about a tenth what you'd pay for the big US models. For a business running lots of AI requests, that gap can mean the difference between an experiment and an affordable product.

Note: pricing is approximate at launch (1 June 2026) and may change.

Catch 1: you couldn't download it at launch. 'Open-weight' is a term that means you can download the AI model and run it on your own computer — no monthly subscription, no company in the middle. Despite using that label, MiniMax launched M3 without the actual downloadable files. They said the files would appear on a model-sharing site called Hugging Face within about ten days. Until that happens, 'open-weight' is a promise rather than a done deal. This matters most if you're a developer who wants to test, modify or self-host the model — for now, you're renting access, like with any other commercial AI.

Catch 2: the impressive scores came from MiniMax. When a company says its new product scores higher than the competition, it's worth knowing who ran the test. In this case, the 59% SWE-Bench Pro figure is MiniMax's own, run on their own infrastructure. That's not unusual — all labs publish their own initial benchmarks — but it does mean independent researchers couldn't check the number at launch (because the downloadable version wasn't available). It may well hold up; you just can't confirm it yet. There's also a weaker area worth knowing about: M3 scores under 12% on a test called ARC-AGI-2, which measures more abstract, flexible reasoning. It's genuinely strong at coding; it's not the all-rounder that headline might imply.

MiniMax is a Shanghai AI company with a genuine track record — their earlier models (the M-series) have been considered competitive by people who've tested them. They're not a newcomer making extravagant first claims. The 'weights in about ten days' commitment is a cadence other Chinese labs have generally kept — so the question is more about verification than credibility. Separately, MiniMax's shares on the Hong Kong stock exchange apparently swung up about 5% on the announcement before falling back lower the same day, which is a useful proxy for 'exciting but unresolved' — the market's two-minute verdict, and probably a reasonable one for you too.