GoKawiilJeff: open-source 0.8B zero-shot decision models compatible with Jev API
An independent developer released Jeff, a set of small fine-tuned models (based on Qwen3.5 and Gemma 4) that return calibrated probabilities across user-defined options in a single forward pass, taking roughly 22-28ms per decision on consumer hardware like an RTX PRO 6000 or Apple M4 Max. The models share Jev's request format but were trained entirely locally using synthetic data generated by an open model, with no cloud GPUs or closed-model training data involved. Three model variants are available on Hugging Face: Jeff-Qwen3.5-0.8B, Jeff-Qwen3.5-2B, and Jeff-Gemma4-E2B.
GoKawiil's interpretation of the reporting above, not reported fact.
By matching Jev's request format while running on local hardware, Jeff could let developers swap in a much cheaper, faster classification model without rewriting integration code, particularly for latency-sensitive tasks like routing or moderation. The developer's own benchmarks suggest fine-tuning on domain-specific examples can dramatically improve accuracy over zero-shot use, which may make Jeff more practical for narrow applications than general-purpose reasoning tasks. Its fully local, open training pipeline also demonstrates that small specialized models can be built without cloud infrastructure or proprietary training data.
- Jeff models run decisions in ~22-28ms per call on consumer-grade or workstation GPUs, with no text generation needed.
- Fine-tuning on custom examples reportedly raised voice-navigation accuracy from 31.7% to 95.8% in under 30 minutes on one GPU.
- Jeff is an independent project, not affiliated with TypeSafe (Jev's maker), though it mirrors Jev's request format and builds on the open-source AutoJev recipe.
Apple MacBook Pro 16-inch M4 Max — This article benchmarks local model inference on an Apple M4 Max via MLX, making the MacBook Pro with M4 Max an ideal machine for running these tiny decision models entirely on-device. Its unified memory and Apple Silicon performance are well suited to fast local ML experimentation like the Jeff models described here.
See Apple MacBook Pro 16-inch M4 Max on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-09-28
Published there as: “Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.