Model
Inference & Serving
Mixtral 8x7B
1Sparse 2mixture-of-experts 3model 4rivaling 5GPT-3.5 6quality
the six Ws · specification
W1
Who
Mistral AI released Mixtral 8x7B, its first sparse mixture-of-experts model.
W2
What
A 47-billion-parameter (12.9B active) MoE language model matching or beating Llama 2 70B and GPT-3.5 on many benchmarks.
W3
Where
Distributed on Hugging Face and via Mistral's API.
W4
When
Mixtral 8x7B was released in December 2023, with Mixtral 8x22B following in April 2024.
W5
Why
It proved sparse MoE architectures could deliver top-tier performance at lower inference cost, popularizing the approach in open models.
W6
With
Runs via vLLM, transformers, or llama.cpp, requiring MoE-aware inference routing.
W7
Watch
Open-weight release by Mistral AI under Apache 2.0 license.
alternatives
works with