Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms

Jeff is a lightweight decision model (0.8B parameters) fine-tuned from Qwen and Gemma models that performs zero-shot classification at high speed (~22-28ms per decision) without generating text or requiring parsing. The model can be trained locally on a single GPU in about 2 hours and easily integrates into code using the same request format as Jev, allowing users to classify inputs into custom categories described in plain language. Fine-tuning on task-specific examples can significantly improve accuracy, as demonstrated by a voice-navigation example that improved from 31.7% to 95.8% accuracy in under 30 minutes.