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. The announcement landed on a Monday morning. By Tuesday, you could call the API. The weights were a different story.
That gap - API live, weights pending - is the right place to start if you want to understand both what is real about M3 and what requires healthy skepticism.
What is actually new
The architecture is the genuinely interesting part. Standard transformer attention has quadratic computational complexity: as context length grows, compute cost grows as the square of the sequence length. This is why most models with long context windows are "technically capable" of using a million tokens but practically painful and expensive to run at that length.
The core architectural innovation in M3 is MiniMax Sparse Attention (MSA). This design uses a lightweight index branch to scan incoming tokens and select which blocks of the key-value cache are relevant to a given input, and the main attention layer then processes only those selected blocks. 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.
There is a telling backstory here. M3 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 and shipping order-of-magnitude speedups is a meaningful self-correction, not a minor point release.
MSA is deliberately different from DeepSeek's Multi-head Latent Attention (MLA), which compresses KV state and trades off some long-context precision - MSA sidesteps that compression-precision tradeoff. Whether that design choice holds up under independent scrutiny is one of the things the technical report, when it ships, should clarify.
The efficiency numbers are the most trustworthy part of this launch. They describe a mechanism. They do not depend on someone's benchmark harness.
The benchmark situation
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.
That number is worth scrutinizing. MiniMax has not disclosed the total parameter count of M3. 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.
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. That is not a scandal - most frontier labs publish benchmark numbers before reviewers can replicate them - but it does mean you are currently trusting a press release, not a peer-reviewed result.
Pricing and what it actually means
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. At promotional rates on OpenRouter, M3 launched at $0.30 per million input tokens and $1.20 per million output tokens during a 50%-off launch promotion, with regular pricing at $0.60/$2.40.
For teams running agentic coding loops where a single task might consume hundreds of thousands of tokens across dozens of tool calls, that price differential is not cosmetic. A model that scores within a few percentage points of the closed frontier at one-tenth the cost deserves a serious trial - on your actual tasks, not on someone else's leaderboard numbers.
The context window matters for the same reason. Most models with long context windows are technically capable of processing a million tokens but practically painful to use at that length. Latency balloons, costs spike, and the economics fall apart before you hit the limit. M3 is the first model that makes a credible case that 1 million tokens could actually be an operational feature rather than a marketing ceiling. Whether that claim holds in practice is worth testing on your own representative input sizes before you design a pipeline around it.
The data residency question
This is the part that often gets waved past in benchmark round-ups, so it is worth being direct.
MiniMax is a Chinese company. China's 2017 National Intelligence Law requires organizations operating in China to cooperate with state intelligence efforts when requested. This is not unique to MiniMax, and it does not make the model unusable. But it is a material consideration for anyone processing sensitive data, proprietary code, regulated financial information, or anything with compliance requirements.
It can be privacy-safe if you self-host the open weights so data never leaves your systems. The hosted API needs more scrutiny because of where the company operates. Run a compliance review either way.
The self-hosting path is where M3's open-weight status earns its keep. Once the weights are on Hugging Face - MiniMax announced that M3 weights would be released within approximately 10 days of the June 1, 2026 launch
- teams with serious data sensitivity can evaluate it without their prompts ever touching MiniMax infrastructure. A teammate like Beagle, which processes Slack and Teams messages, would need that deployment model before this model belongs anywhere near production internal conversations.
The licensing terms for M3 had not been published at launch, and M3 may follow a similar approach. Check the license before you build on it.
Incremental or genuinely new?
The honest answer is both, in different parts.
The MSA architecture is a substantive engineering contribution. Abandoning sparse attention, watching the infrastructure catch up, and returning to it with production-grade speedups is the kind of thing that should survive independent verification. If the technical report matches the launch claims, this is a real architectural advance.
The benchmark numbers are closer to marketing until independent runs reproduce them. 59.0% on SWE-Bench Pro is plausible given the trajectory of open-weight models - Kimi K2.6 held the highest SWE-Bench Pro score of any open-weight model at its April 2026 release at 58.6% , so M3 claiming 59.0% six weeks later is not implausible. But "not implausible" is not the same as verified.
The combination - frontier-adjacent coding, a genuinely usable million-token context, native multimodality, and open weights under a permissive-enough license - would be a meaningful moment for the open-weight ecosystem if it holds up. Watch for the first independent evals. Run a narrow pilot on your own workloads. Do not design a production system around the vendor's benchmark table alone.