How much does a Llama prompt consume?

Source: Table 4 of Jegham et al., arXiv 2505.09598 (version 6, 24 November 2025). The energy figures are the table's, in Wh, mean per request. "n/a": the study did not test these two models on a long prompt (context-window limit). They are modelled estimates for an assumed AWS hosting, not measurements taken on the servers.

How much does a Llama prompt use, depending on model size?

Between 0.05 Wh and 2.23 Wh on a short prompt, and between 0.44 Wh and 25.20 Wh on a long one. The study tests ten Llama versions and three prompt sizes: short (100 tokens in, 300 out), medium (1,000 in and 1,000 out) and long (10,000 in, 1,500 out). A token is a chunk of a word. The table above gives all twenty-eight values from the study. From Llama 3.1 8B to Llama 3.1 405B, consumption is multiplied by about 43 on a short prompt and by about 57 on a long one. This site computes those ratios: the study does not publish them.

Does a bigger Llama always use more?

No. Size matters a lot, but it does not explain everything. Llama 3.3 70B uses 0.24 Wh on a short prompt and Llama 3.1 70B uses 1.27 Wh: 5.4 times more, for the same stated size. On a long prompt the gap grows to 13.3 times (1.45 Wh against 19.18 Wh). Llama 3.2 Vision 90B, though bigger, also uses less than Llama 3.1 70B on a long prompt: 6.85 Wh against 19.18 Wh. The study does not explain these gaps. Its method starts from the response time observed on public services: a version that answers faster weighs less, whatever its size.

Do these figures apply to a Llama you run yourself?

Not directly. The study assumes Llama runs at AWS, on Nvidia servers of the H100 or H200 class, with the average efficiency of AWS data centers. A Llama installed on your own machines, at another host or on another graphics card will use a different amount. The same study finds large gaps between two hosts for the same model: see this site's DeepSeek page.

What about water and CO₂?

The study details them only in a chart, except for the smallest models. The text states that Llama 3.2 1B, Llama 3.2 3B, Llama 3.1 8B and Llama 3 8B emit less than 0.3 g of CO₂e and use less than 4 mL of water, even on a long prompt. This site therefore gives no water or CO₂ value for the large Llama models.

Are these figures reliable, and do they hold over time?

They give an order of magnitude and ratios between versions, not an exact bill. This is an arXiv preprint that models consumption without access to the servers, and it has been revised: any comparison with a figure read elsewhere should name the version, here the November 24, 2025 one. The study also publishes a standard deviation for each value, which this page's table does not reproduce. A short prompt measures the cost of a simple question, not that of an average use.