MiniMax M3: The First Open-Weight Model to Enter Frontier Territory
Over the past two years I watched the same line being drawn again and again. On one side: the proprietary frontier — first-class coding, a million-token context window, native multimodality, all in a single model. On the other: the open-weight world, which always seemed to have two of three things but never all three at once. You could get a solid open coding model, or one with long context, or one that was multimodal. Wanting all three meant signing an API contract with someone in San Francisco.
On June 1st, MiniMax launched M3 and erased that line — at least on paper.
What was actually announced
M3 launched simultaneously via MiniMax Code, the company’s subscription token plan, and its API. The pitch is straightforward: it’s the first open-weight model that combines frontier-level coding, a 1M token context window, and native image and video understanding in a single architecture. MiniMax frames these three capabilities as the bare minimum of any closed frontier model today — and claims M3 is the only open-weight model that brings all three together.
The benchmark that will grab the headlines is SWE-Bench Pro, where M3 reports 59.0%. According to MiniMax’s own numbers, that puts it ahead of GPT-4.5 and Gemini 3.1 Pro, and just behind Opus 4.7. The rest of the coding and agents suite is in the same ballpark: 66.0% on Terminal-Bench 2.1, 74.2% on MCP Atlas, and a reported advantage over Opus 4.7 in autonomous web search (BrowseComp).
The pricing is where things get interesting for anyone running this at scale: $0.60 per million input tokens and $2.40 per million output tokens. That’s not a rounding error compared to proprietary frontier pricing — it’s a different cost category altogether.
The part that really matters: MSA
This is where I’d tell a technical colleague to slow down and read carefully. A 1M token context window is no longer novel — several models advertise one. The question that decides whether you’ll ever actually use it is: how much does it cost per token when the window is full?
This is the real announcement. M3 introduces a new attention architecture that MiniMax calls MSA (MiniMax Sparse Attention). Instead of comparing every token against all others — the quadratic cost trap of full attention — MSA divides the stored context into blocks, runs a pre-filtering pass to decide which blocks are actually relevant to the current query, and only performs the full computation on those.
The numbers MiniMax reports: with 1M tokens, computation per token drops to approximately 1/20th of the previous generation. More than 9× faster prefill, more than 15× faster decode. In other words, the million-token window isn’t a spec sheet trophy you’ll avoid touching because of the bill — it’s designed to be economically usable.
If those numbers hold up in independent testing, this is the genuinely important contribution. Benchmark rankings will shift every couple of weeks; an attention mechanism that cheapens long context is the kind of thing that changes how you architect a system.
The autonomy tests are worth looking at
MiniMax ran three internal long-horizon experiments that go beyond single-prompt benchmarks. The one that stuck with me: they gave M3 a paper that won an award at ICLR 2025 on fine-tuning LLMs and asked it to independently reproduce the work. M3 ran autonomously for almost twelve hours, produced 18 commits and 23 experimental figures, and reproduced the paper’s core findings.
In a second test, they asked M3 to optimize an FP8 GEMM kernel on NVIDIA Hopper GPUs from a non-functional skeleton with no reference implementation — a task that, according to MiniMax, takes an experienced team one to two weeks. Over about 24 hours it ran 147 benchmark submissions and pushed hardware utilization from 7.6% to 71.3%, not reaching its best solution until attempt 145.
Whether or not those exact numbers survive scrutiny, the framing tells you where the field is heading: the next generation of coding agents won’t be judged on single-shot code generation, but on long-duration collaboration, planning, and the ability to push through plateaus without giving up.
Where skepticism is warranted
I’m not going to hand you a press release dressed up as analysis. Some things to keep in mind before reorganizing your stack around this:
These are MiniMax’s own benchmarks. Every number above comes from the company’s launch materials. The SWE-Bench Pro numbers in particular have a habit of looking different once independent labs run their own harnesses.
The weights haven’t dropped yet. “Open-weight” is the whole pitch — but at launch time, MiniMax only promised the weights and a technical report on Hugging Face and GitHub within roughly ten days. Until they appear and someone outside MiniMax can run the model, the open-weight claim is a commitment, not a fact you can build on.
The frontier has already moved. The framing of “just behind Opus 4.7” was true on June 1st. Anthropic has since launched Opus 4.8. Comparing a freshly launched model to last month’s frontier is standard marketing practice, and you should adjust your mental ranking accordingly.
And yes — M3 comes from a Chinese lab, which for some teams carries data governance and procurement implications worth thinking through before adopting. I mention it because it’s a real factor in enterprise decisions, not because it changes the technical picture.
My read
Strip away the launch day noise and two things stand. First: if the open weights drop as promised and the benchmarks survive independent replication, M3 is the first time the open-weight world has credibly claimed all three frontier capabilities at once — and that’s a structural change, not an incremental one. Second: MSA is the part I’d actually be watching. Cheap long context changes what you can build; a benchmark ranking that’s true this week won’t necessarily be next week.
For now, the honest position is: promising enough to test, too early to bet on. Drop it in a sandbox when the weights come out, run it on your own workloads with your own harness, and decide from there. But don’t let the 59% headline think for you.