Jevstiller is an open‑source project that builds a tiny local model to answer routine Jev queries on your own hardware. The local model learns from Jev responses and only forwards uncertain requests to Jev’s cloud service. Its creators say it can answer familiar queries in about 15 ms and avoid the $42 per billion token fee. They claim the tool matches Jev’s answers 98 % of the time while keeping the rest of the traffic in the cloud for safety.
The system continuously audits a fixed 2 % of requests against Jev to keep the agreement rate in check and retrains itself when the rate drops.
Why it matters
Running Jev locally can slash AI usage costs and speed up response times for common tasks.
