Skip to content
Longterm Wiki

Llama 4 Maverick

Meta AI (FAIR)Open Weight

Llama 4 Maverick was released April 5, 2025 as a 17B active parameter mixture-of-experts model (400B total, 128 experts). Scored 92.2% on MMLU, outperforming GPT-4o and Gemini 2.0 Flash. Featured a 1M token context window and native multimodal support. Could be deployed on a single H100 DGX host.

Developer
Meta AI (FAIR)
Released
2025-04-05
Context Window
1M tokens

Benchmarks6

Knowledge

2 benchmarks
92.2%#3/37
93th percentile
80.5%#2/15
90th percentile

Reasoning

1 benchmark
69.8%#15/34
57th percentile

Coding

1 benchmark
43.4%#7/9
28th percentile

Multimodal

2 benchmarks
73.4%#2/3
50th percentile
94.4%#1/2
50th percentile
Percentile among tested models:Top 25%50-75%25-50%Bottom 25%

Llama Family5

ModelTierReleasedInput $/MTok
Llama 4 Scout2025-04-05
Llama 3.32024-12-06
Llama 3.12024-07-23
Llama 32024-04-18
Llama 22023-07-18

Details

Model FamilyLlama
Generation4
Release Date2025-04-05
Parameters400B
Context Window1M tokens
Open WeightYes
Modalitytext, image

Capabilities2

tool-usevision

Sources1

Tags

llamametaopen-weightmixture-of-expertsfrontier-ai