MiniMax M3

A 427B MoE that runs one million tokens of context and eats video. MiniMax M3 is the largest open-weight release to come out of the minimax-community license, and the specs are singular: 60 layers with 128 routed experts plus one shared expert (4+1 active per token), GQA attention at 4 key-value heads, a 200,064-token vocabulary, and MiniMax Sparse Attention - the in-house operator that makes the 1,048,576-token context usable, with claimed 9x prefill and 15x decode at 1M context versus their M2 generation. It reads images and video natively: the pipeline tag is image-text-to-text. 168k downloads and 1.5k likes on Hugging Face so far.

What it is for. Long-horizon agent work over huge mixed corpora - repos with video, docs, screenshots, code - where the 1M context plus multimodal input does the job in one pass. The vendor-recommended servers are SGLang, vLLM, KTransformers, unsloth, and ROCm’s ATOM stack for MXFP4, which is the honest preview of where MXFP4 support lives first in the open ecosystem.

Fit on this site’s tracked hardware. The unsloth Dynamic quants put UD-Q4_K_XL at 264.9GB and UD-Q2_K_XL at 143.0GB; even the smallest quality-usable run (UD-IQ1_M = 128.4GB from the file sums) does not fit our workstation tier with KV headroom, so M3 is a premier-class box: 256GB pooled Macs, 2x DGX Spark clusters, or the 384GB Spark build. Anything smaller should run M2.7 instead.

Open weights under the minimax-community license - commercial use permitted below 100M monthly active users per the license text, the same structure as the M2 generations. Downloaded checkpoints are the real thing: no gating, no phone-home.