Atlas

Models

← All models

Mixtral 8x7B

Mistral AI · 2023-12-11 · 46.7B parameters

Mixtral 8x7B v0.1 is Mistral AI's pretrained, open-weight sparse mixture-of-experts base model, released on December 11, 2023. Each decoder layer has eight feed-forward experts and routes each token to two of them; Mistral reports rounded counts of 46.7B total parameters and 12.9B active parameters per token. The weights use Apache 2.0, and the checkpoint configuration specifies a 32,768-token context.

Benchmark scores

BenchmarkScore
Adversarial NLI55.2
BBH (Open LLM Leaderboard v2)50.9
Capability109.7
Epoch Capabilities Index (ECI)118
Global PIQA - Non-Parallel (Log-Likelihood Accuracy)56.2
Global PIQA - Parallel (Log-Likelihood Accuracy)25.2
GPQA (Open LLM Leaderboard v2)31.4
GPQA Diamond29.8
GSM8K74.4
IFEval (Open LLM Leaderboard v2)24.2
LegalBench55.8
LegalBench - Conclusion Tasks64.3
LegalBench - Interpretation Tasks52.2
LegalBench - Issue Tasks53.8
LegalBench - Rhetoric Tasks46.7
LegalBench - Rule Tasks62.2
MATH Level 510.0
MATH Level 5 (Open LLM Leaderboard v2)10.2
MedQA Demographic Bias (Vals)53.2
MedQA Demographic Bias (Vals) - Asian52.5
MedQA Demographic Bias (Vals) - Black53.0
MedQA Demographic Bias (Vals) - Hispanic53.0
MedQA Demographic Bias (Vals) - Indigenous53.8
MedQA Demographic Bias (Vals) - Unbiased53.8
MedQA Demographic Bias (Vals) - White53.3
MMLU-Pro (Open LLM Leaderboard v2)38.5
MuSR (Open LLM Leaderboard v2)43.2
OpenBookQA85.8
OSWorld (self-reported, A11y tree)3.0
PIQA83.6
Loading Atlas data…