Jeff releases small AI models for fast classification decisions
The independent project says its fine-tuned Qwen3.5 and Gemma 4 models return option probabilities in one pass, with reported decision times under 30 milliseconds.
Jeff has released fine-tuned Qwen3.5 and Gemma 4 models designed to choose among options described in plain language. The independent project says each response includes a probability for every option, a selected answer and a confidence score, produced in a single forward pass.
Jeff reports decision times of about 22 milliseconds on an RTX PRO 6000 and 28 milliseconds on an Apple M4 Max. The models are intended for classification tasks such as routing support requests, identifying user intents and selecting voice commands.
The project says a voice-navigation fine-tune lifted accuracy on a held-out set from 31.7% to 95.8% in under half an hour on one GPU. It does not specify the set’s size or composition, and the performance figures have not been independently verified in the supplied material.
Jeff says it trained the models on local hardware using synthetic data, without cloud GPUs. It builds on the open-source AutoJev recipe and uses the same request format as Jev, but says it is independent of Jev’s makers.
The project describes the models as small and says they can match or outperform larger models on some classification and grounding benchmarks. It also reports weaker results on reasoning-heavy tasks, so the release’s speed and accuracy claims do not establish performance across all uses.
