Featherless AI

Simple Jev is an open source library from Featherless AI that turns any Hugging Face model into a structured classification endpoint, returning typed choices, rubric scores, and truth or support judgments as JSON instead of free-form text. It targets developers who want the ergonomics of TypeSafe's Jev, a System One-style model that emits structured, type-safe decisions with calibrated confidence rather than generating prose , but without being locked into a closed model or provider. The launch matters because it opens up that interaction pattern to the wider ecosystem of open weights, and adds vision-capable models to the mix, two gaps the team highlights in the launch tweet. Under the hood, the server evaluates the exact common token prefix once and reuses that prefix's KV cache across batches of question suffixes, then reads the next-token logits for the allowed answer labels to construct the JSON response . This avoids generating and parsing a prose or JSON answer token by token, and reusing the context can reduce repeated computation when several questions refer to the same input . The project ships with a hosted playground and a no-auth demo API at simple-jev.featherless.ai for developers who want to try it against sample models before self-hosting or wiring it into Featherless infrastructure. The library is built and shipped by Featherless AI, co-founded by Eugene Cheah, Harrison Vanderbyl, and Wesley George, the same core team behind RWKV , the Linux Foundation attention-free model architecture project. Eugene (@picocreator) is posting the launch, and the code lives at github.com/featherless-ai/simple-jev alongside guides for dataset preparation, teacher labeling, LoRA training, and export back to the HF-based server so teams can distill their own small task-specific classifiers.