MiniMax M3 lands with frontier coding numbers — but no weights to inspect
MiniMax's June 1 launch claims 59.0% on SWE-Bench Pro and a 1M-token context at roughly one-twentieth the compute cost of M2.7. The catch: weights, architecture detail, and final license are promised within ten days, not on day one.
The interesting piece in MiniMax's M3 launch isn't the headline benchmark — it's the gap between the marketing and the artifact. MiniMax announced the model June 1, 2026 with vendor-run scores of 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, and 83.5 on BrowseComp, numbers it says edge past GPT-5.5 and Gemini 3.1 Pro on agentic coding while closing in on Claude Opus 4.7. The catch is that the open weights, training code, inference operators, and final license terms have not been published; MiniMax says they'll appear on Hugging Face and GitHub within ten days.
M3 is built on MiniMax's proprietary Sparse Attention (MSA) architecture and ships with a 1M-token context window — five times the 200K window of M2.7. The company claims MSA cuts per-token compute at 1M context to roughly one-twentieth of the previous generation, with more than 9x faster prefill and more than 15x faster decoding. Hosted inference is being priced at 5-10% of the cost of comparable proprietary frontier models, which is the lever MiniMax is using to seed adoption while the weights paperwork catches up.
That gap matters because the open-weights market has built two years of trust on a same-day publish rhythm. Until M3 ships safetensors that independent engineers can load, the "open-weight" label is a company commitment rather than a verifiable property — buyers can't fine-tune on private data, sovereign clouds can't deploy in a VPC, and security teams can't audit the architecture for the edge cases the benchmarks don't cover. The launch lands the same day as NVIDIA's Nemotron 3 Ultra, which shipped weights, technical report, and license on day one — a contrast the open-source community is going to keep pointing at.
VentureBeat — MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost → · The Decoder — MiniMax M3: Open-weight model with a million-token context challenges proprietary leaders → · TechTimes — MiniMax M3 Open-Weight Coding Model: Frontier Claims, Unverified Benchmarks →