MiniMax unveils M3, an open-weights model touting coding-agentic gains and 1M context
In a thread on X, MiniMax cites 59.0% on SWE-Bench Pro and a new sparse attention scheme that scales context to 1M, with an API promo rolling out this week.
By Ryan Merket · · updated
Why it matters
If the reported numbers hold, M3 pushes open-weights models further into practical coding and agent workflows while adding 1M context via sparse attention. That combination could let teams self-host or fine-tune a more capable agentic stack without closed-model lock-in. A short-term API discount signals MiniMax is angling for fast developer adoption, turning benchmark wins into early integrations.
MiniMax announced M3 in a thread on X, casting it as the first open-weights model to combine strong coding and agentic performance with a 1M context window via its MiniMax Sparse Attention.
On headline benchmarks, MiniMax reports 59.0% on SWE-Bench Pro, 66.0% on Terminal Bench 2.1, 34.8% on SWE-fficiency, 28.8% on KernelBench Hard, and 74.2% on MCP Atlas. MiniMax says M3 scales context to 1M using a sparse attention approach and points to more details in a link.
MiniMax also outlined launch pricing: 50% off standard usage (<=512K context) for the first 7 days, with priority access available via [email protected] and self-serve access for all users coming in the next few days. The company shared promotion terms alongside the announcement.
For a limited time you can use the new model for free on OpenCode:
