LLaMA 4 Maverick Environmental Impact

Energy Consumption

LLaMA 4 Maverick activates 17B of its 400B total parameters per token, routing across 128 experts. Despite the same 17B active parameter count as Scout, Maverick requires an 8×H100 node (with FP8 quantisation) because all 400B parameters must be loaded in memory. This larger memory footprint and more complex routing increase energy per query to approximately 1.8 Wh.

Power Source & Carbon

As an open-source model, Maverick is deployed on diverse infrastructure. It requires a multi-GPU setup (8×H100), limiting it to cloud or enterprise hardware. Meta's own data centres run about 60% on renewable energy.

Water Usage

At approximately 3.4 mL per query when run in a data centre. The multi-GPU requirement means more cooling is needed than for single-GPU models like Scout.

How much energy does LLaMA 4 Maverick use per query?

Each LLaMA 4 Maverick query consumes approximately 1.8 Wh of energy. This is 6x more than a traditional Google search (~0.3 Wh).

What is LLaMA 4 Maverick's carbon footprint?

Based on the carbon intensity of Self-hosted (varies), each query produces approximately 0.72 g of CO2. The grid in this region has a carbon intensity of 475 g CO2/kWh with 27% renewable energy.

How much water does LLaMA 4 Maverick use?

Each query consumes approximately 3 mL of water, primarily used for cooling the data centers that process the request.

How does LLaMA 4 Maverick compare to a Google search?

A LLaMA 4 Maverick query uses 6x more than a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while LLaMA 4 Maverick uses 1.8 Wh.

Architecture

Multimodal Transformer (MoE, 128 experts, iRoPE)

Parameters

400B

Context window

1,000,000 tokens

Release date

2025-04-05

Open source

Yes

Training data cutoff

2025-02