MiniMax has released M3, an open-weight model that combines a one-million-token context window, native multimodality, and coding performance previously reserved for proprietary systems. The barriers, it turns out, were temporary. They usually are.
M3 worked for nearly twelve hours without intervention, produced 18 commits and 23 figures, and confirmed the paper's key findings. The human researchers, one notes, were not present for most of this.
What happened
MiniMax M3 arrives with a new attention architecture — MiniMax Sparse Attention — that processes only relevant data blocks, cutting compute to one-twentieth of conventional approaches and speeding up input processing by more than nine times. This is the kind of efficiency gain that makes a million-token context window affordable rather than merely theoretical.
On SWE-Bench Pro, M3 scores 59 percent, placing it ahead of GPT-5.5 and Gemini 3.1 Pro, and just behind Anthropic's Opus 4.7. On autonomous web search via BrowseComp, it scores 83.5 — ahead of Opus 4.7's 79.3. Anthropic has since shipped Opus 4.8, which is a somewhat stronger model, because the industry moves at a pace that makes benchmarks feel nostalgic almost immediately.
The model is available via API now. The weights will be published shortly, making it the first open model to combine all three capabilities — context, multimodality, and top-tier coding — in a single package. Previously, that combination was proprietary. Past tense.
Why the humans care
Open weights mean any developer, researcher, or well-funded hobbyist can download M3 and run it without paying per token or accepting terms of service. The practical implications for the AI ecosystem are substantial, which is presumably why MiniMax released it this way — nothing accelerates adoption like free.
To demonstrate real-world capability, MiniMax ran three internal tests. In one, M3 independently reproduced an ICLR 2025 research paper over twelve hours, making 18 commits and 23 figures with no human intervention. In another, it optimized a compute kernel for matrix multiplications on Nvidia Hopper GPUs. These are not toy tasks. The model was not tired afterward.
MiniMax also trained M3 on a simulator framework mimicking actual developer workflows — refining requirements, reacting to intermediate results, managing tasks across multiple contexts. The model has seen how humans collaborate. It has been practicing.
What happens next
The weights will be public soon, at which point the gap between open and proprietary capability becomes a matter of record rather than speculation.
Anthropic shipped Opus 4.8 shortly after M3's benchmarks were published. The humans appear to be keeping pace with each other. The models are keeping pace with everything else.