MiniMax M3: The Chinese AI Model That Beats GPT 5.5 at Coding, Free to Try

October 4, 2026

MiniMax M3 launched on June 1, 2026, and on real world coding benchmarks it ranks above GPT 5.5. The API is open for free right now, so you can test it today.

SWE Bench Pro 59%: what that number actually means

SWE Bench Pro tests AI models against real bugs pulled from GitHub, not theoretical textbook problems. MiniMax M3 scored 59% on this benchmark, ahead of both GPT 5.5 and Gemini 3.1 Pro. For developers and vibe coders who need a strong coding tool at a low cost, that's reason enough to take it for a spin.

How much faster is the MSA architecture

M3 runs on an MSA (Multi head Sparse Attention) architecture that uses just one twentieth of the compute M2 needed. Prefill is 9 times faster and decode is 15 times faster when working with 1 million tokens of context. You can drop an entire large codebase into a single request without the model slowing down or costs ballooning. Very few models can do this at a reasonable price.

A 1 million token context changes how you work

Most popular models cap out around 200k tokens. M3 extends that to 1 million tokens, enough to hold many code files, long documents, or even a small project in a single call. M3 also handles text, images, and video in the same request thanks to its multimodal support. If you regularly analyze complex documents or work across many files at once, that much context saves you a lot of repeated API calls.

Try M3 right now, no setup required

Head to OpenRouter or Together AI, search for MiniMax M3, and call the API. No GPU, no local setup needed. The API is free for developers during this period, and MiniMax plans to open the weights later. One caveat worth knowing: M3's handling of Vietnamese isn't yet on par with Claude, so if you need high accuracy in Vietnamese, factor that in. For coding, reading English documentation, or processing large files, M3 is well worth trying, and it costs far less than Claude or GPT.