MiniMax: M3

Open-weight, natively multimodal, 1M-token context on the new MiniMax Sparse Attention architecture: ~1/20th per-token compute at 1M context and 59.0% on SWE-Bench Pro, ahead of GPT-5.5 in MiniMax's own comparisons. All numbers vendor-run, and the weights were promised "within about 10 days" of the announcement. The open-weight release as a futures contract.

Sources: official announcement, MarkTechPost, TechTimes (benchmark caveats)