MiniMax M3 reaches 59 percent on SWE-Bench Pro and runs fastest on SambaCloud

Chinese company MiniMax has released M3, an open-weights model aimed squarely at long-horizon agents and complex software-engineering tasks. The model is available now to developers on SambaCloud, where it runs faster than on any other platform, and it replaces M2.7 as the strongest model in MiniMax’s portfolio.

The core architectural advance is MiniMax Sparse Attention (MSA), a new attention mechanism that cuts per-token compute to one-twentieth of the previous generation. The result is a one-million-token context window with over 9× prefill acceleration and more than 15× decoding acceleration while retaining performance comparable to full attention on most capabilities. In practice, entire codebases, long papers and multi-day agent logs fit into a single context window without truncation.

On coding and agent benchmarks M3 scores 59% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1 and 74.2% on MCP Atlas, placing it among the strongest options in the open-weights category. It also posts 34.8% on SWE-fficiency and 28.8% on KernelBench Hard. MiniMax says training incorporated an interactive user simulator that mimics how real developers collaborate — clarifying requirements, fixing solutions and iterating across multiple turns — so the model functions as an active partner rather than a passive instruction follower.

In internal tests M3 ran autonomously for roughly 24 hours to optimise an FP8 GEMM CUDA kernel. During the run it made 147 benchmark submissions and 1,959 tool calls, and peak hardware utilisation jumped from 7.6% to 71.3%, a 9.4× improvement, with no human intervention. In a separate experiment the model independently reproduced the core experiments of an ICLR 2025 outstanding paper, producing 18 commits and 23 experimental figures over about 12 hours.

M3 was trained on text, image and video from step 0, not as a modality bolted on later. This lets it read charts, formulas and screenshots, and operate a desktop computer — a capability that opens the door to computer-use agents and end-to-end workflows across applications, files and systems. MiniMax also notes strong results in real office workflows including search, office tasks and tool orchestration, with emerging capability in the financial domain.