The claim sounds like every other launch-week press release: a new open-weight model that beats GPT-5.5 on a coding benchmark and costs a tenth as much. Shanghai-based MiniMax launched M3 on June 1, 2026, positioning it as the first open-weight system to combine frontier coding-agent performance, a one-million-token context window, and native multimodal capabilities - including image, video, and desktop computer operation - in a single model. That is a real list of things, and most models only do one of them well. The question is whether M3 actually delivers or just claims to.
Let's start with what is genuinely new before getting to the caveats.
The architecture is not incremental
The most interesting thing about M3 is not the benchmark number - it is how MiniMax got the 1M-token context window to be economically viable.
The technical centerpiece is MiniMax Sparse Attention, or MSA - a new attention architecture designed to make one-million-token context windows economically viable for production use. Standard transformer models process attention across every token in a context window, a calculation that grows quadratically as context length increases. MSA replaces that with a two-stage mechanism: a lightweight index branch first scans incoming tokens and selects which blocks of the key-value cache are actually relevant, then runs the expensive attention computation only on those selected blocks.
Unlike DeepSeek's Multi-head Latent Attention (MLA), which compresses keys and values into a low-dimensional latent space, MSA operates on a standard GQA backbone but utilizes block-level selection on real, uncompressed Key-Values. The practical consequence: at 1-million-token context length, MSA reduces per-token compute to one-twentieth of the prior generation, delivers more than 9x faster prefill and more than 15x faster decoding.
M3 represents a notable shift in two specific areas. First, it demonstrates that sparse attention can work at production scale for long-context models. MiniMax itself abandoned sparse attention during its entire M2 generation in favor of full attention, calling the infrastructure "not yet mature" at the time. Returning to sparse attention with MSA and achieving order-of-magnitude speedups suggests the technology has caught up.
That is a real architectural bet that paid off - or appears to have, pending the technical report.
The multimodal angle is also substantive rather than bolt-on. Unlike models that add vision capabilities after text pretraining, MiniMax says M3 was trained on interleaved text-image sequences from "Step 0," building multimodal understanding into the base model rather than adding it as a fine-tuning layer.
What the benchmarks actually show
On SWE-Bench Pro, an established software development benchmark, M3 scores 59 percent according to MiniMax. That puts it ahead of GPT-5.5 and Gemini 3.1 Pro, but just behind Opus 4.7. On longer autonomous tasks, MiniMax ran M3 against a GPU kernel optimization problem: the model was given only a task description, a benchmark script, and a non-functional code skeleton with no reference solution. After about 24 hours, it had pushed Hopper hardware utilization from 7.6 to 71.3 percent. Most other tested models gave up after a few dozen attempts, while M3 worked through several plateaus and didn't reach its best solution until attempt 145.
Those are compelling demos. But there is a catch that matters.
Several benchmark results were obtained on MiniMax's own infrastructure using agent scaffolding such as Claude Code and Mini-SWE-Agent. Independent third-party verification is still pending, and M3 has not yet appeared on the DeepSWE board for long-horizon software tasks. A model that scores 59% with proprietary scaffolding on its home infrastructure may land somewhere different when third parties run the same eval cold.
On PostTrainBench, a research autonomy evaluation, M3 scored 0.37, trailing Opus 4.7 at 0.42 and GPT-5.5 at 0.39. The model is not uniformly ahead of its closed-source peers - it is selectively ahead on the benchmarks MiniMax chose to publish.
The "open-weight" designation is a commitment, not yet a fact
At the time of writing, at the time of launch, neither the weights nor the technical report had been released. MiniMax said both would be made available within ten days of launch, targeting publication on Hugging Face and GitHub for private cluster deployment and fine-tuning. That means developers cannot yet inspect the architecture details, verify the training setup, assess the safety behavior under edge cases, or confirm the licensing terms. Until the weights ship and independent engineers can reproduce the architecture claims, M3's open-weight designation is a company commitment - not a verifiable fact.
The licensing situation adds another layer. MiniMax-M2 shipped under a modified-MIT license, but the newer M2.7 license restricts commercial use of the model or derivatives without prior written authorization. If M3 follows that precedent, expect downloadable weights with a non-commercial default and enterprise licensing available through direct sales.
Three things to watch: the open weights and technical report landing on Hugging Face within the promised ten days, the first independent benchmark runs that strip away MiniMax's own scaffolding, and M3's eventual appearance on neutral boards like DeepSWE.
The price is real, though
M3 pairs a genuinely novel sparse attention architecture with frontier-adjacent benchmark scores at a price point well below Western closed-source competitors. The cost gap is real: M3's launch pricing is roughly one-tenth the input cost of Claude Opus 4.7 and GPT-5.5, a difference that compounds materially in agentic workflows.
Output speed runs at approximately 100 tokens per second, roughly 3x faster than Claude Opus.
For teams running high-volume coding agent loops, that arithmetic is hard to ignore - even before self-hosting enters the picture.
The benchmark scores come from MiniMax itself. Test it on your own work before you trust it in production. An AI teammate like Beagle can help you structure that evaluation in Slack or Teams - capturing results and surfacing comparisons without losing context between sessions.
The data sovereignty question is not optional to skip
If your team handles regulated or sensitive data, the API option requires a careful read before you route anything through it. Developers deciding whether to route their coding workflows through M3 need to weigh three things: benchmark scores are company-reported and run on MiniMax's own infrastructure; promised open weights have not been released; and China's 2017 National Intelligence Law requires MiniMax to "support, assist, and cooperate" with Chinese government intelligence work, an obligation that applies to every prompt processed through the company's API endpoint, regardless of where the user is located.
The self-hosted path, once the weights are confirmed and the license is clear, sidesteps this entirely. Self-hosting the open weights means data never leaves your systems. The hosted API needs more scrutiny because of where the company operates. Run a compliance review either way.
So, can it actually match Claude on coding?
On MiniMax's own evals: close, but not quite. On neutral evals: we do not know yet. The MSA architecture is a real contribution - sparse attention at this scale, applied cleanly to production context windows, is not trivial engineering. The agentic persistence demos (12 hours autonomously reproducing a research paper, 145 iterations on a GPU kernel) are at least worth taking seriously as signals of something different in the training approach.
The honest answer right now is: wait two weeks. Watch for the weights to land on Hugging Face, check the license, and look for the first independent SWE-Bench runs that are not run on MiniMax infrastructure. If those hold up, M3 will be the most interesting open-weight coding model available. If they don't, it will join the long list of models that benchmarked better at home than anywhere else.