Solo developer ships Jeff, an open decision‑model family
Jeff mimics the request format of the proprietary Jev API but runs locally in about 22 to 28 milliseconds, trained on synthetic data with no closed-model outputs, its GitHub README says.
A developer released Jeff, a family of three open fine-tuned models, according to its GitHub README. Jeff returns calibrated probabilities for multiple-choice decisions from a single forward pass. The models are based on Qwen3.5 and Gemma 4 and range from 1.7 to 9.3 gigabytes. They are released under the Apache 2.0 licence, with MIT-licensed code.
Jeff uses the same request format as Jev, a proprietary decision-routing API, the README says. It is not affiliated with Jev's maker, TypeSafe, and was built instead on the open AutoJev recipe. The project reports inference times of about 22 milliseconds on an Nvidia RTX Pro 6000 and 28 milliseconds on an Apple M4 Max. It puts Jev's own published times at roughly 114 to 212 milliseconds.
The models were fine-tuned on synthetic data from Qwen3.8-Flash-Next and trained on local hardware, the README says. No outputs from closed commercial models were used in training, it adds. The repository had drawn 81 stars and four forks by the time it reached the Hacker News front page with 98 points and 11 comments.
The claims about speed and training data are the developer's own, published in the repository rather than verified by a third party. The README does not disclose accuracy figures alongside its speed numbers.