jeff

Jeff is a set of small open models that answer multiple-choice, yes-or-no and scoring questions with a probability per option instead of generated text. It speaks the same request format as TypeSafe's closed Jev API, so it drops in for classification, routing and guardrail checks. Code is MIT, weights are Apache 2.0, and the training code builds on AutoJev from Denis Yarats of Perplexity.

There are three models: 0.8B and 2B versions built on Qwen3.5, plus a heavier Gemma 4 variant. The small one needs 1.7GB and answers in about 22 milliseconds on a workstation GPU, 28 on an M4 Max, or about half a second on a big CPU. Setup is uv, a Hugging Face download and jeff-serve, which listens on localhost by default with optional key auth. Training code is included, and the author reports a voice-command fine-tune jumping from 32% to 96% accuracy in half an hour on one GPU.

Teams sorting tickets, intents or sentiment at volume should self-host it instead of paying Jev's $42 per billion input tokens. Its benchmark average lands just under Jev's published number, and an outside user re-ran the suite and matched the author's scores. NandhaKishorM/laya is smaller and covers 100+ languages; jaredpalmer/kev scales up to 27B.

The catch: the Jev comparison uses Jev's published figures on a different sample, not side-by-side runs, and on reasoning-heavy tests like BBH and WinoGrande Jeff trails badly. It's English and text only. The latest version also trained on two popular intent benchmarks, so those scores are no longer zero-shot.