Atlas

Models

← All models

Llama 4 Scout Instruct

Meta · 2025-04-05 · 108.6B parameters

Llama 4 Scout Instruct is Meta's instruction-tuned, natively multimodal mixture-of-experts checkpoint with 16 routed experts plus a shared expert and native text-and-image input. Meta describes Scout as 17B activated and 109B total parameters (both rounded), with roughly 40T pretraining tokens and an August 2024 data cutoff; the official artifact configuration supports up to 10,485,760 tokens.

Benchmark scores

BenchmarkScore
AA-LCR25.8
AA-Omniscience Index-52.4
AidanBench334
AIME 2024 (avg@8)27.9
AIME 2024-202519.0
AIME 202514.0
AIME 2025 (avg@8)10.0
ARC-AGI-10.5
ARC-AGI-20.0
Arena Text — Business, Management & Financial Ops (No Style Control)1258
Arena Text — Chinese (No Style Control)1254
Arena Text — Creative Writing (No Style Control)1250
Arena Text — English (No Style Control)1303
Arena Text — Entertainment, Sports & Media (No Style Control)1244
Arena Text — Exclude Ties (No Style Control)1211
Arena Text — Exclude Ties (Style Controlled)1270
Arena Text — Expert (No Style Control)1238
Arena Text — French (No Style Control)1289
Arena Text — French (Style Controlled)1333
Arena Text — German (No Style Control)1272
Arena Text — Hard Prompts (No Style Control)1268
Arena Text — Hard Prompts English (No Style Control)1292
Arena Text — Instruction Following (No Style Control)1251
Arena Text — Japanese (No Style Control)1210
Arena Text — Korean (No Style Control)1213
Arena Text — Legal & Government (No Style Control)1286
Arena Text — Life, Physical & Social Science (No Style Control)1280
Arena Text — Longer Query (No Style Control)1268
Arena Text — Medicine & Healthcare (No Style Control)1275
Arena Text — Multi-turn (No Style Control)1282
Loading Atlas data…