Local AI Matches 89% of Queries

Researchers evaluated 20 local language models and eight hardware accelerators in the United States. They measured how accurately these models answer over one million real-world questions. The team introduced “intelligence per watt” (IPW) as a way to measure both accuracy and power efficiency.

Results show local language models correctly answer roughly 89% of the queries tested. Analysis from 2023 to 2025 indicates IPW improved by a factor of 5.3. This improvement comes from better algorithms and hardware.

Local accelerators used about 1.4 times less power than cloud accelerators when running the same models. The study acknowledges performance varies depending on the question’s subject matter. The researchers continue to track IPW as a key metric for shifting demand away from centralized cloud infrastructure.