MiniMax-M3 BF16 safetensors

MiniMax's natively multimodal MoE model (428B total, ~23B active), released June 2026 and trained mixed-modality from the first step. MiniMax Sparse Attention delivers roughly 9× prefill and 15× decode speedups over M2 at 1M-token context. Three reasoning modes: enabled, adaptive, disabled.

License note: the MiniMax Community License permits redistribution, with free non-commercial use and commercial use for organizations under $20M annual revenue — read the terms for anything bigger.

This torrent mirrors the official Hugging Face release weights.

Swarm verified via DHT metadata fetch on 2026-08-08; size and file count are taken from the torrent metadata itself. Created and permanently seeded by drforbin.ai — this is the site's own release of the official Hugging Face weights.