Mixtral 8x7B Original release weights

Mistral AI's sparse mixture-of-experts model, released 8 December 2023 as a magnet link on X ("mixtral-8x7b-32kseqlen"). 46.7B total parameters with ~12.9B active per token, 32K context, Apache-2.0.

This is the raw release checkpoint exactly as Mistral published it. The Hugging Face-format conversion lives on the linked model card.

Swarm verified via DHT metadata fetch on 2026-07-23; size and file count are taken from the torrent metadata itself.