← Back to directory
Mistral AI · Released 2024-04

Mixtral 8x22B

MoE classic from 2024 — 8 experts of 22B, 39B active per token. Apache-2.0 makes it the go-to open MoE for production teams.

Apache-2.0Commercial use OKgeneral
Params (max)
141B
Variants
141B / 47B
Context window
64K tokens
MMLU
77.8
HumanEval
75.3
GSM8K
88.4
Min VRAM (fp16, smallest variant)
80GB
Smallest Q4 GGUF
~26GB
Languages supported
5
Pros
  • ✓Truly free license
  • ✓Sparse efficiency
  • ✓Battle-tested
Cons
  • ×Older — newer models are smarter
  • ×MoE memory footprint big

Highlights

  • ●Apache-2.0
  • ●39B active params
  • ●Mature in production

Where to download

Hugging Face: mistralai/Mixtral-8x22B-Instruct-v0.1
Or via Ollama (ollama pull mixtral-8x22b) or LM Studio's in-app browser.
Homepage: https://mistral.ai