MiniMax M3

Announced June 1, 2026 with weights following on Hugging Face, MiniMax M3 folds frontier coding, million-token context, and native multimodality into one open checkpoint of about 428B total parameters with 23B active. Its defining piece is MiniMax Sparse Attention, an architecture that replaces full attention with learned sparse patterns and delivers a claimed 9x prefill and 15x decode speedup over the previous M2 at 1M-token lengths. The model was trained on mixed modalities from the first step, so image and video understanding are native rather than bolted on; published results include 59.0 on SWE-Bench Pro and 66.0 on Terminal-Bench 2.1. Local deployment works through vLLM, SGLang, Transformers, KTransformers, and Unsloth with recommended sampling of temperature 1.0 and top_p 0.95, though the size still demands a multi-GPU node. Weights are distributed under the MiniMax Community license, with the company's API and Token Plan subscriptions covering hosted use.