MiniMax just dropped M3 with 1M context and desktop control

MiniMax M3 is a new open-weight model built specifically for frontier coding and complex agentic workflows, packing a 1M token context window, native multimodality, and desktop computer control into a single architecture.

What it can do

On software engineering benchmarks, M3 hits 59.0% on SWE-Bench Pro, outpacing frontier closed-source options like GPT-5.5 and Gemini 3.1 Pro. It also clocks 66.0% on Terminal-Bench 2.1 and 74.2% on MCP Atlas.

The real highlight, though, is its long-horizon autonomous execution. In internal evaluations, M3 took an ICLR 2025 award-winning research paper and reproduced its core experiments from scratch over nearly 12 hours of uninterrupted work, logging 18 commits and 23 figures without human intervention.

On a grueling systems engineering benchmark, it was tasked with optimizing an FP8 GEMM kernel on NVIDIA Hopper GPUs. Given only an unrunnable Triton skeleton, M3 ground away for 24 hours, making 147 benchmark runs and nearly 2,000 tool calls. It pushed peak hardware utilization from 7.6% all the way to 71.3%, achieving a 9.4x speedup. Notably, its best solution arrived on the 145th attempt, continuing to explore long after most other frontier models gave up.

Combined with native image and video understanding, it can also pilot desktop environments to complete multi-step office and terminal tasks.

How it works and how to try it

To make a 1M context usable in production, MiniMax developed MSA (MiniMax Sparse Attention). Instead of traditional full attention that chokes on quadratic complexity, MSA uses clean KV-block pre-filtering with contiguous memory reads.

At the 1M token mark, per-token compute drops to just 1/20 of the previous generation. This architecture delivers over 9x faster prefilling and 15x faster decoding than M2, easily beating open-source sparse alternatives like Flash-Sparse-Attention in arithmetic intensity.

MiniMax also trained M3 on multimodal interleaved data right from step zero, pairing it with an interactive developer simulator to train multi-turn collaboration patterns rather than simple one-shot prompt answering.

The weights are open, and developers can immediately take it for a spin through MiniMax Code, the Token Plan, and official API endpoints.

Why it matters

For developers building local or self-hosted agents, M3 removes the painful tradeoff between open weights and frontier agentic muscle. Until now, pairing 1M context with native vision and persistent multi-hour problem-solving meant paying steep closed-API tolls.

We think M3's standout trait is its stubbornness during long tasks. Sticking through dozens of plateaued submissions to land an optimized CUDA kernel shows genuine architectural stamina. If you need a capable code partner for heavy refactoring or autonomous multi-agent pipelines, M3 should be on your immediate test list.