Overview
Llama 4 Maverick is Meta's flagship open-weight model and its first to use a mixture-of-experts design. It packs 400B total parameters across 128 experts but activates just 17B per token, so it punches well above its compute weight — beating GPT-4o and Gemini 2.0 Flash on the LMSYS Chatbot Arena at launch with a 1417 Elo. It's natively multimodal (text and image) with a 1M-token context window.
Maverick's appeal is open-weight frontier capability you can host yourself: strong general reasoning, MMLU-Pro around 80%, and competitive coding, all under Meta's community licence. The launch drew some benchmark-transparency criticism, but as a self-hostable multimodal model at this scale, Maverick remains a cornerstone of the open ecosystem.
Key capabilities
- 400B MoE with only 17B active per token
- Natively multimodal (text + image)
- 1M-token context window
- Beat GPT-4o and Gemini 2.0 Flash on Arena at launch
At a glance
LMSYS Arena
1417 Elo (launch)
MMLU-Pro
~80.5%
Context
1M tokens
Pricing: Low — open weights, efficient MoE inference.
Pros & cons
What we like
- Frontier-adjacent open multimodal model
- Efficient MoE keeps inference affordable
- Large context and broad capability
Trade-offs
- Newer rivals have overtaken its launch scores
- Benchmark-transparency questions at launch
The verdict
Llama 4 Maverick brought efficient mixture-of-experts and native multimodality to Meta's open lineup, and it's still a strong, self-hostable flagship. For teams that want frontier-adjacent capability under an open licence, it's a foundational choice.
Best for: Self-hosted, multimodal, frontier-adjacent open AI.
Frequently asked questions
How is Llama 4 Maverick different from Scout?
Maverick is the larger 400B flagship (128 experts) tuned for top quality; Scout is a lighter 109B model (16 experts) optimised for speed and an enormous 10M-token context that fits on a single H100.
Is Llama 4 Maverick free to use?
It's released under Meta's community licence with open weights, so you can self-host it — subject to the licence's terms and usage thresholds.
