Solo developer's open‑source Jev clone runs on a laptop CPU and has 1,200 stars
Federico Elia's jevos answers yes-or-no questions in 26 to 112 milliseconds on a CPU with no GPU, and his launch post for it has about 2,400 likes on X.
Federico Elia, who posts as FederCr, released jevos on GitHub this week as an open-source alternative to TypeSafe AI's Jev. His X post calls it the fastest model ever made and says it is ten times faster than Jev. The post had about 2,400 likes when we read it on 1 October, and the repository shows roughly 1,200 stars and 130 forks.
https://x.com/FederCr/status/2105675061990391959
Jev is a closed-weight cloud model that returns typed decisions such as yes, no or a score instead of generated text. Elia's argument is that the network round trip makes Jev slow and that every call is billed per token. jevos runs on the user's own machine, according to its README, in a single binary with no Python or GPU needed.
The README says jevos is built on MiniCPM, uses 8-bit weights through OpenVINO and needs about 1 gigabyte of memory. It lists latencies of 26 to 112 milliseconds on a CPU against 344 and 345 milliseconds for Jev, and an accuracy of 80.3 percent on yes-or-no questions. Those numbers come from the author, and we have not rerun them.
The repository is MIT-licensed, handles multiple-choice questions and returns confidence scores and probabilities. It also ships GGUF files for llama.cpp. Its context window is 8,192 tokens, far shorter than the 64,000 Cloudflare lists for its new Clef models.
jevos is one of a crowd. Von, a 395-million-parameter ModernBERT model, claims under 15 milliseconds, and several Gemma-based replicas exist, according to listings on GitHub. Jev launched in early access on 15 September, so a small ecosystem of clones has formed in about two weeks.