मिनीमॅक्स M3

MiniMax's native multimodal model with a million-token context, built with roughly 428B total and 23B activated parameters. It uses MiniMax Sparse Attention for efficient long context and is trained on mixed modalities from inception, so text, images and video are integrated semantically rather than having vision bolted on afterwards. MiniMax presents it as the first open-weight model with frontier coding, agentic reasoning and native multimodality at once.

MiniMax makes a combination claim rather than a peak-score one. Frontier coding, a million tokens of context and native multimodality are each unremarkable in a closed frontier model, and M3 is the lab's argument that one set of open weights can carry all three at the same time. What makes that affordable is MSA, the sparse attention design the team proposes: it partitions the key-value cache into blocks more precisely than the alternatives it names, and its operator uses the blocks as the outer loop so each is read exactly once with contiguous memory access.

The efficiency figures are what turn the context from a number into something usable. MiniMax reports the operator running more than four times faster than two open sparse-attention implementations, per-token compute at a million tokens of one twentieth of the previous generation, and end-to-end speedups of more than nine times in prefill and more than fifteen times in decode. Across several ablations it says MSA matched full attention on the vast majority of capabilities, which is the claim that matters most and the easiest one to get wrong.

For coding the lab reports 59.0% on SWE-bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, 28.8% on the hard tier of a GPU kernel benchmark and 74.2% on MCP Atlas. More interesting than the numbers is the training argument underneath them. MiniMax says most code-agent training and evaluation assumes a single-turn task, which is not how anyone actually works, so it built an interactive user simulator that reproduces requirement elaboration, mid-task correction, task switching and multi-round iteration, and trained and evaluated against that instead.

Its long-run examples follow the same shape. Handed an award-winning machine learning paper and asked to reproduce it independently, M3 ran for close to twelve hours, produced 18 commits and 23 experimental figures and completed the core experiments, including reproducing the effect the paper is known for and verifying the mitigation it proposes. On an FP8 matrix-multiplication kernel with no reference implementation available to imitate, it ran for roughly 24 hours across 147 benchmark submissions and 1,959 tool calls, lifting hardware peak utilisation from 7.6% to 71.3%. MiniMax notes that most other models it tried stopped making progress within the first 30 submissions, while M3's best result arrived on submission 145, after several plateaus it worked through rather than gave up on.

The lab is explicit about where it is not yet ahead. On the post-training benchmark, where the model must autonomously synthesise data, train, evaluate and iterate on four base models inside twelve hours, it scored 0.37, behind Claude Opus 4.7 at 0.42 and GPT-5.5 at 0.39 though clearly ahead of everything else it tested. It describes its agentic performance in the financial domain as only beginning to be usable rather than solved. Weights and a technical report were promised within ten days of the post, so this is a model you could host yourself; the copy served here is MiniMax's.