MiniMax M3: Open-Weight Frontier Model With 1M-Token Context Beats GPT-5.5 on Coding at a Fraction of the Cost
Chinese AI lab MiniMax released M3 on June 1, combining frontier-level coding, a 1-million-token context window, and native multimodality in a single open-weight model — scoring 59% on SWE-Bench Pro while cutting per-token compute costs by 20x.
MiniMax M3 launched on June 1, 2026, and the headline is straightforward: it’s the first open-weight model to simultaneously deliver frontier-level coding, a 1-million-token context window, and native multimodality. The pitch turned out to be real. M3 scores 59.0% on SWE-Bench Pro — beating GPT-5.5 and Gemini 3.1 Pro, and approaching Claude Opus 4.7. It does this while using one-twentieth the per-token compute of the prior MiniMax generation at 1M context.
The MSA Architecture
The efficiency story comes from MiniMax Sparse Attention (MSA), an architecture specifically designed to handle very long contexts without the quadratic cost of standard full attention. MSA delivers more than 9x faster prefill and more than 15x faster decoding at 1 million tokens compared to the previous architecture. That’s the difference between 1M context being a theoretical capability and a practical one.
For reference: running GPT-4-class models at 1M context with full attention is prohibitively expensive for most use cases. MSA makes it tractable at production scale.
What M3 Actually Does
Three things in combination, which no other open-weight model has previously offered together:
-
Frontier coding: 59.0% on SWE-Bench Pro puts it ahead of GPT-5.5 and Gemini 3.1 Pro on that benchmark. That’s not cherry-picked — SWE-Bench Pro is the current gold standard for real-world software engineering task completion.
-
1M-token context: Enough to load multiple large codebases, full research documents, or long conversation histories without chunking. Most open-weight models cap at 128K–256K tokens.
-
Native multimodality: M3 handles text, images, and video natively in the same model, without a separate vision encoder pipeline.
Availability and Open Weights
M3 is live on the MiniMax API and through MiniMax Code subscriptions. Open weights were committed at launch, with MiniMax targeting publication on Hugging Face and GitHub approximately ten days after the June 1 release — putting the open-weight drop around June 11.
Until then, independent verification of the benchmark numbers isn’t possible. MiniMax ran the evaluations on their own infrastructure with agent scaffolding, which is standard practice but also the condition under which benchmark inflation most often occurs. The 59.0% SWE-Bench Pro number deserves scrutiny once the weights are public.
Cost Profile
MiniMax prices M3 competitively against Tier-2 proprietary models, not against GPT-5.5 or Claude Opus pricing. Exact per-token pricing varies by plan, but the company’s positioning is explicit: approximately 5–10% of the per-call cost of competing frontier proprietary models at scale. For teams building on top of LLMs, that’s the gap between a feature being economically viable and not.
Why It Matters for Developers
The open-weight commitment is the important part. Proprietary frontier models can be changed, rate-limited, or repriced by their owners at any time. An open-weight model at 59% SWE-Bench Pro that you can self-host or fine-tune for your domain is a qualitatively different offering — especially with a 1M-token context that doesn’t require chunking strategies for large codebases.
If the numbers hold up after independent verification on June 11, M3 will be the default recommendation for teams that need frontier coding capability without a hard dependency on a proprietary API.