What matters in AI.

Subscribe

Compute

A test of 86 open-weight models shows large differences in energy use

Thinking mode uses about 3.7 times more energy on average. FP8 uses about one quarter less energy with a small change in score.

Energy use, relative to 1× Ranking of energy multiples from 86 open-weight model checkpoints: worst gap 27×, thinking on 3.7×, typical gap 2.6 to 4.8×, equal-scoring best 1×, FP8 0.75×. Score held, energy did not. Energy use, relative to 1× 1× Worst gap Thinking on Typical gap (median) Equal-scoring best FP8 27× 3.7× 2.6–4.8× 1× 0.75× about a quarter less Score held. Energy did not. 86 checkpoints · 7 benchmarks · 6 platforms
Energy use compared with a baseline of 1×, from 86 open-weight model checkpoints tested on seven benchmarks and six hardware platforms.

Sources

  1. Measuring the Inference Energy Efficiency of Open-Weight LLMs across Benchmarks, Reasoning Modes, and Hardwarearxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·14 OCT2026

Posted

Tags

Learn the terms in this story

Guides for this story

More in Compute

All Compute news