Mixtral 8x7B Instruct v0.1

Permissive

Mistral AI · Mistral · Released 2023-12

8x7B (46.7B total, ~12.9B active)Text

The model that put sparse mixture-of-experts on the open-weight map: 70B-class quality at 13B-class inference cost, no license strings attached.

Strengths

  • +Apache 2.0, unrestricted commercial use
  • +Strong quality-to-active-compute ratio via MoE
  • +Good multilingual and code performance

Limitations

  • -Total memory footprint (~47B params) still needs 24-48GB VRAM even though only ~13B is active per token
  • -MoE routing adds serving complexity vs dense models

License

Apache 2.0

use commercially with attribution niceties

Hardware

wants 24-48GB of VRAM - a 3090/4090-class card or better

Links

Stats

mistralai/Mixtral-8x7B-Instruct-v0.1— downloads·— likes

via Hugging Face

Related models