DeepSeek · released 2025-12-01
DeepSeek-V3.2 uses about 0.057 Wh – 1.06 Wh of electricity to answer a typical question. That places it 181 of 266 among the models we can measure, and every figure here is calculated the same way as every other model.
The capability figure of 146.19 is a composite over 58 benchmarks — Epoch AI's Capabilities Index, which folds in software engineering, agentic tool use, long-context comprehension and creative writing as well as science reasoning. It is one number standing for a lot of separate tests, so treat a small gap between two models as noise rather than a ranking.
8,905 people have voted on it in blind head-to-head comparisons on LMArena, giving it a rating of 1425. This is the only figure here that measures whether anyone actually prefers the answers, rather than how the model scores on a test. Broken down, people rate it 1453 on coding, 1403 on creative writing, 1413 on following instructions.
Inferred The maker publishes no size at all, so it is inferred from how capable the model is and when it was released. A wide bracket, not a figure. Every model on this site is calculated the same way, so the comparison holds even where the absolute value carries uncertainty. Our figures are good to roughly a factor of two to three, which is why we only ever state a difference of 5× or more.
Energy cannot be worked out without knowing how big a model is. Of the 15 models from DeepSeek that we track, 10 can be assessed at all — the rest publish too little for anyone outside the company to calculate anything.
Capability and the roster come from Epoch AI, whose Capabilities Index is the 58-benchmark composite quoted above. Energy is calculated with EcoLogits and calibrated against laboratory measurements from ML.ENERGY. Human preference comes from LMArena. All four are public and free; the full method is here.
These figures are published under CC BY 4.0 — use them anywhere, including commercially, with credit to The AI Footprint.
Compare DeepSeek-V3.2 with another model