MiniMax-M3

What MiniMax-M3 is

MiniMax-M3 is a mixture-of-experts (MoE) language model released by MiniMax in June 2026. Its weights were posted to Hugging Face on 2 June 2026, and the accompanying paper went up on arXiv as 2606.13392 on 11 June ^[2][3][1]. The model has about 428B total parameters, of which roughly 23B are activated for each token it processes ^[1]. The model card is not perfectly consistent on this point: the safetensors metadata lists 427B parameters, while the card text rounds the figure to about 428B ^[1].

In a mixture-of-experts model, each token is routed through a small subset of the network instead of the whole thing, so the compute spent per token is far lower than the total parameter count would suggest. That is the trade M3 makes: large capacity stored in the weights, a small slice of it used at any one moment.

Sparse attention for long contexts

The main architectural change is MiniMax Sparse Attention (MSA), which MiniMax designed to replace GQA-style attention ^[1]. Attention is the part of a transformer that lets each token weigh other tokens in the context. Grouped-query attention, the scheme MSA replaces, shares key and value heads across query heads to save memory, but its cost still grows with the length of the context. Sparse attention instead computes only part of the token-to-token comparisons, which is what makes very long inputs affordable.

M3 supports a context of 1M tokens ^[1]. At that length MiniMax reported prefill 9 times faster and decoding 15 times faster than the earlier M2, with per-token compute cut to one-twentieth ^[1]. Prefill is the pass that reads the prompt before generation starts, and decoding is the token-by-token generation that follows; speeding up both is what makes a context this large usable in practice rather than merely nominally supported.

Natively multimodal

M3 was trained on text, images and video from the start, rather than having vision bolted on after the fact ^[1]. Training on video in particular means the model saw sequences with strong temporal structure during pretraining, which fits the long-context emphasis of the release.

Reported results

MiniMax published scores on several standard evaluations: SWE-bench Verified 80.5, SWE-bench Pro 59, MMMU-Pro 78.1 and APEX-Agents 27.7 ^[1]. SWE-bench asks a model to resolve real software issues drawn from open repositories, and the Pro variant uses a harder task set; MMMU-Pro tests reasoning over images, diagrams and other visual material; APEX-Agents measures performance on multi-step agentic tasks. The model card also lists Long-Horizon-Terminal-Bench at 38.5 and Lhtb2 Solved at 46.9, both reported as a mean reward multiplied by 100 ^[1]. Those two are long-horizon terminal benchmarks, which test a model across extended sequences of command-line steps rather than a single answer.

Licensing and what followed

The weights were released under a minimax-community license ^[1]. A follow-up, MiniMax-M3.1-Flash, arrived in October 2026 ^[3].

Where it sits

The release is best read as a bid to make very long context practical, not just possible. The combination of sparse attention, a 1M-token window, native multimodality and software-engineering and agent benchmark reporting points at workloads where the model reads a large amount of material and then acts on it over many steps. The parameter count also keeps M3 in the sparse, selectively activated camp: large enough to hold broad knowledge, cheap enough per token to serve long inputs.