Jevstiller: open source tool distills Jev for local inference

An open source project called Jevstiller distills Jev's outputs into a small local model that handles familiar requests on-device. Local answers come in as little as 15 ms instead of ~300 ms and spend no Jev tokens, while uncertain queries are forwarded upstream; the team claims 98% agreement with Jev.

• Local answers in 15 ms versus ~300 ms on Jev
• No token spend at $42 per billion input tokens
• A fixed 2% of requests always go to Jev for auditing
• In a soak test local share fell from 90% to 9% in 4 minutes after Jev changed answers