Atlas

Models

← All models

Mixtral 8x22B

Mistral AI · 2024-04-17 · 140.6B parameters

Mixtral 8x22B v0.1 is Mistral AI's pretrained base sparse Mixture-of-Experts text model, released on April 17, 2024. Its configuration routes each token through two of eight local experts, supports 65,536-token sequences, and uses grouped-query attention; Mistral describes the model at rounded precision as 141B total parameters with 39B active. The weights are released under Apache 2.0.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)62.4
Capability117.7
Epoch Capabilities Index (ECI)121
FORTRESS Average Risk Score (ARS)56.1
GPQA (Open LLM Leaderboard v2)37.6
GPQA Diamond34.1
IFEval (Open LLM Leaderboard v2)25.8
MATH Level 524.2
MATH Level 5 (Open LLM Leaderboard v2)18.4
MMLU77.8
MMLU-Pro (Open LLM Leaderboard v2)46.4
MuSR (Open LLM Leaderboard v2)40.4
Loading Atlas data…