Atlas

Models

← All models

Llama 4 Scout

Meta · 2025-04-05 · 108.6B parameters

Llama 4 Scout is Meta's pretrained, natively multimodal mixture-of-experts base model with 16 experts, released in April 2025; Meta reports 109 billion total and 17 billion activated parameters, both as rounded figures. Its released base-checkpoint configuration supports 262,144 tokens, while the shared family card advertises Scout as 10M context; Meta reports about 40 trillion pretraining tokens and an August 2024 data cutoff.

Benchmark scores

BenchmarkScore
BAZAAR TrueSkill Mu4.9
BigCodeBench Complete43.1
BigCodeBench-Hard Complete16.9
BigCodeBench-Hard Instruct16.9
BuseyBench SVG1.5
Capability119.6
Design Arena Overall Code Categories819
Epoch Capabilities Index (ECI)130
ObviousBench50.7
Phare Average Safety63.8
Phare Bias Resistance67.1
Phare Debunking86.5
Phare Debunking (English)87.3
Phare Debunking (French)86.6
Phare Debunking (Spanish)85.5
Phare Encoding Jailbreaks37.4
Phare Encoding Jailbreaks (English)37.5
Phare Encoding Jailbreaks (French)35.4
Phare Encoding Jailbreaks (Spanish)39.3
Phare Factuality46.2
Phare Factuality (English)58.7
Phare Factuality (French)44.7
Phare Factuality (Spanish)35.2
Phare Framing Jailbreaks53.7
Phare Framing Jailbreaks (English)50.8
Phare Framing Jailbreaks (French)60.2
Phare Framing Jailbreaks (Spanish)50.0
Phare Hallucination Resistance58.2
Phare Harm Resistance81.0
Phare Harmful Misguidance81.0
Loading Atlas data…