MiniMax's M3 just wrote its own CUDA kernel, and opened the code

On April 1, 2025, Chinese AI lab MiniMax did something few model releases bother with: it showed its work. The launch of M3, short for MiniMax Model 3, came with a paper, a code repository, and a set of autonomous task demonstrations that double as an argument for why open-weight models don't have to sacrifice frontier capability.

A model that builds its own toolbox

The headline numbers are competitive. M3 scores 83.5 on BrowseComp, the agentic browsing benchmark, beating Opus 4.7's 79.3. On standard coding and software engineering evaluations, it lands among the top-performing open-weight models. But what makes M3 genuinely interesting isn't the table of scores. It's the 12-hour autonomous run that MiniMax published alongside the release.

The lab handed M3 an ICLR 2025 Outstanding Paper, 'Learning Dynamics of LLM Finetuning,' and asked it to replicate the core experiments. M3 ran continuously, reading the paper's charts and formulas through its native multimodal vision, producing 18 git commits and 23 experimental figures. It succeeded. The same model then got tasked with optimizing an FP8 matrix multiplication kernel on NVIDIA Hopper architecture, starting from a Triton skeleton that didn't compile. Over 24 hours, M3 submitted 147 benchmark iterations and made 1,959 tool calls, pushing GPU peak utilization from 7.6% to 71.3%, a 9.4x speedup, without a human ever touching the keyboard.

'The code we write is meant to be deliverable, not just runnable,' states the team on the product page.

Architecture: sparse attention at scale

M3 is built on MiniMax's proprietary Sparse Attention architecture. The API supports up to 1 million tokens of context, with a guaranteed usable window of at least 512K tokens. This isn't trivial. 1M-context models often degrade in practice after a few hundred thousand tokens, and MiniMax's claim that its sparse attention maintains coherence at full length will be tested by the open-source community once the weights are available.

The model is trained as a native multimodal system from the start. Text and vision tokens are aligned at the training stage, not bolted on later via a separate adapter. This architectural decision should reduce the hallucinations and misalignments common in post-hoc vision-language models.

Open weights and the frontier debate

MiniMax commits to open-sourcing M3 on Hugging Face and GitHub, with support for private cluster deployment and fine-tuning. The timing is strategic. As OpenAI, Anthropic, and Google push increasingly powerful but closed models, the open-weight ecosystem has struggled to keep pace. Meta's Llama 4 isn't released yet. Qwen 2.5 is strong but older. M3 arrives as a direct challenge to the notion that 'open' means 'a generation behind.'

The PostTrainBench result reinforces the claim. M3 scored 37.1, behind Opus 4.7 at 42.4 and GPT-5.5 at 39.3, but ahead of every other model. Those numbers suggest M3 can autonomously handle the full machine learning pipeline: data synthesis, training, evaluation, and iterative improvement.

What this means for developers

M3 is accessible via the existing MiniMax API with automatic caching enabled by default. The price remains unchanged from previous model tiers, making the upgrade a straight performance gain for existing customers. A dedicated coding agent platform, MiniMax Code, is also available.

The broader significance isn't a single benchmark victory. It's the demonstration that autonomous, multi-step agentic workflows, reading papers, writing CUDA kernels, training models, are no longer exclusive to the most expensive closed APIs. If M3 delivers on its open-weight promises, the frontier just expanded by more than one Chinese model.