jev-align
Teach a Jev classifier your labels by reviewing only its least certain answers
Emerged finds new products worth a look, daily — what’s new today
If you classify thousands of documents by asking a small model several questions about each one, Flide reads each document once and answers all the questions in one pass.
Flide is a self-hosted inference server for small decision models in the Jev family. When many questions target the same input, it processes that context once, then forks the model state so each question gets exactly the answer it would get alone. The team reports up to 9.5x lower latency than vLLM and SGLang on QuALITY, and near parity for large models asked one question. Apache-2.0 and on PyPI. Linux and NVIDIA only; about 6 GB of VRAM runs the 0.8B model.
More like this
Jev Ultrafast
An open browser agent that finishes a flight search in about seven seconds
Jinfer
An AI inference engine that runs inside the JVM, with no Python sidecar
jevals
Run dozens of agent evals per trace in one sub-second request, no judge prompt
Kev
Open 0.8B to 9B decision models for choice, yes/no and rating questions
Laya
Small encoder models that answer routing and triage questions in one pass