Llama 4 Scout is Meta's efficient multimodal language model with 16 experts activating 17 billion parameters out of 109B total. It supports native multimodal input (text and image) across 12 languages with a 10-million-token context window — one of the longest available — and uses early fusion for seamless modality integration. Designed for high efficiency and local or commercial deployment, it is instruction-tuned for multilingual chat, captioning, and image understanding tasks, released under the Llama 4 Community License.
API|Vision|Open ModelLlama
Knowledge Cutoff
2024-08-31
Input → Output Format
Context Memory
328KIN16KOUT
AI Performance Evaluation
Arena Overall Score
1322
±5As of 2026-05-01
Overall Rank
No.200
30,314 Votes
Arena by Ability
Hard Prompts
1329±6No.198
Expert Knowledge
1308±16No.194
Instruction Following
1299±7No.208
Conversation Memory
1320±9No.195
Creative
1289±10No.197
Coding
1361±9No.198
Math
1309±13No.185
Arena by Occupation
Creative Writing
1306±8No.195
Social Sciences
1336±9No.201
Media
1290±9No.189
Business
1319±9No.195
Healthcare
1342±15No.188
Legal
1344±14No.181
Software
1350±7No.202
Mathematics
1313±14No.185
Source:Arena Intelligence
Overall
AA Intelligence Index
14%↓26%
ForecastBench
54%↓5%
Reasoning & Math
AA Math Index
14%↓60%
GPQA Diamond
59%↓23%
HLE
4.3%↓13%
MMLU-Pro
75%↓6%
AIME 2025
14%↓60%
MATH-500
84%↓9%
Coding
AA Coding Index
6.7%↓30%
LiveCodeBench
30%↓36%
TAU2
16%↓65%
TerminalBench
1.5%↓33%
SciCode
17%↓25%
Language & Instructions
IFBench
40%↓24%
AA-LCR
26%↓36%
Hallucination (HHEM)
7.7%↓2%
Factual (HHEM)
92%↑2%
Output Speed
Standard Mode
128tok/s↑51
First Output 0.57s
