DIY Jev

The "Jev thingy" is gaining significant attention, though some perceive it as overhyped. Performance metrics for different Qwen3 models were shared, measured on a laptop with a 5090 24GB GPU. Qwen3-4B achieved 65.0% accuracy at approximately 27 req/s, Qwen3 27B reached 75.3% at around 2.9 req/s, and Qwen3.6 35B-A3B performed at 75.5% accuracy with about 5 req/s.

This report uniquely details a DIY Rust web server that runs local GGufs from HF in a jev-compatible style, unlike other jev discussions.