MiniMax M3: Open-Weights 1M-Context Agentic Coding Leader
MiniMax has unveiled M3, an open-source agentic-coding model with a 1-million token context window, multimodal support (image, video, and desktop actions), and top-of-leaderboard performance on SWE-bench-Pro — outperforming GPT-5.5. The model uses MiniMax sparse attention to make 1M-context inference practical. API access is available now at approximately $0.20 per million tokens; open weights release in 1–2 weeks.
Why It Matters
M3 combines SOTA agentic-coding performance with an open-weights commitment and pricing roughly 50× cheaper than frontier closed alternatives — a direct challenge to the cost argument that keeps enterprises on closed-model APIs for long-horizon coding workflows.