Llama 4 Maverick is Meta's high-capacity multimodal language model built on a Mixture-of-Experts architecture with 128 experts and 17 billion active parameters per forward pass (400B total). It supports multilingual text and image input across 12 languages with a 1-million-token context window, and uses early fusion for native multimodality. Instruction-tuned for assistant-like interaction and image reasoning, Maverick is optimized for vision-language tasks requiring advanced multimodal understanding and high throughput.
API|Vision|Open ModelLlama
Knowledge Cutoff
2024-08-31
Input → Output Format
Context Memory
1.0MIN16KOUT
AI Performance Evaluation
Arena Overall Score
1327
±4As of 2026-05-01
Overall Rank
No.194
39,997 Votes
Arena by Ability
Hard Prompts
1338±6No.189
Expert Knowledge
1324±13No.178
Instruction Following
1313±6No.191
Conversation Memory
1324±8No.187
Creative
1307±9No.173
Coding
1373±8No.183
Math
1319±11No.175
Arena by Occupation
Creative Writing
1311±7No.183
Social Sciences
1343±8No.190
Media
1295±8No.184
Business
1320±8No.193
Healthcare
1346±13No.181
Legal
1343±12No.183
Software
1359±6No.187
Mathematics
1320±11No.178
Source:Arena Intelligence
Overall
AA Intelligence Index
18%↓21%
ForecastBench
57%↓2%
Reasoning & Math
AA Math Index
19%↓55%
GPQA Diamond
67%↓15%
HLE
4.8%↓13%
MMLU-Pro
81%↓1%
AIME 2025
19%↓55%
MATH-500
89%↓4%
Coding
AA Coding Index
16%↓21%
LiveCodeBench
40%↓26%
TAU2
18%↓63%
TerminalBench
6.8%↓27%
SciCode
33%↓9%
Language & Instructions
IFBench
43%↓20%
AA-LCR
46%↓16%
Hallucination (HHEM)
8.2%↓2%
Factual (HHEM)
92%↑2%
Output Speed
Standard Mode
111tok/s↑34
First Output 0.64s
