MiniMax · 2026-06-01 · major
MiniMax M3 — Open-Weight Frontier Coding, 1M-Token Context, and Native Multimodality on a New Sparse-Attention Architecture
MiniMax shipped M3, an open-weight successor to M2.7 built on a new MiniMax Sparse Attention design. It scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, and serves a 1M-token context window.
M3 packages frontier coding, a million-token context, and native multimodality into one open-weight model.
Key specs
| Context window | 1M tokens |
|---|---|
| Swe bench pro | 59.0% |
| Terminal bench 2.1 | 66.0% |
| Browse comp | 83.5 |
| Plus plan | ~1.7B tokens / $20/mo |
| Max plan | ~5.1B tokens / $50/mo |
| Ultra plan | ~9.8B tokens / $120/mo |
What is it?
M3 is the next model in the MiniMax line after M2.7. It targets coding and agentic work but is natively multimodal — image, video, and desktop operation are trained in from step zero. MiniMax calls it the first open-weight model that combines frontier coding scores, a 1M-token context window, and multimodal input in a single checkpoint.
How does it work?
M3 uses MiniMax Sparse Attention (MSA), a new attention design that pairs softmax expressiveness with a top-k block selection step, keeping sub-quadratic complexity at long contexts. MiniMax reports >9x prefill speedup and >15x decoding speedup at 1M-token context versus M2, with roughly 1/20th the per-token compute. The API, MiniMax Code, and Token Plans go live today; the technical report and open weights are due on Hugging Face and GitHub within about ten days.
Why does it matter?
Open-weight frontier coding models with real 1M-token context have been thin on the ground, and most run text-only. If the weights ship as advertised, M3 gives self-hosters a single model that handles long-repo coding, computer use, and multimodal grounding without stitching together specialised components.
Who is it for?
Open-weight users, agentic-coding builders, long-context power users
Try it
platform.minimax.io — Plus plan: ~1.7B tokens / $20/mo