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

You can now run fast, accurate classification models locally and offline, without cloud dependencies or high costs.

The "Jeff" project, an independent initiative, has released small, Jev-compatible decision models that can be trained locally. These models, fine-tuned from Qwen3.5 and Gemma 4, are designed for zero-shot classification, meaning they can categorize situations based on described options without needing those options to be present in the training data. The 0.8B parameter model, for instance, can make decisions in approximately 22ms on an RTX PRO 6000 GPU and 28ms on an Apple M4 Max. The project emphasizes that these models excel at making quick, calibrated judgment calls and can be easily integrated into local code. While their reasoning capabilities do not match larger models like Jev, their accuracy can be significantly improved with short, local fine-tuning on custom examples. Training data was generated synthetically using an open model, with closed models used only for spot-checking.

The development of "Jeff" signifies a push towards more accessible and auditable AI tooling, directly addressing the growing demand for on-device or local processing capabilities in AI applications. By enabling training on local hardware and utilizing synthetic data from open models, the project champions transparency and cost-efficiency, potentially lowering the barrier to entry for sophisticated AI deployments. This approach challenges the cloud-centric model for AI development and inference, offering a viable alternative for developers prioritizing data privacy and reduced operational expenses.

This release impacts developers building applications that require real-time decision-making without relying on external APIs or cloud infrastructure. Such applications range from interactive gaming and voice command systems to customer support routing and content moderation. The availability of these small, performant models encourages experimentation with AI-driven features in resource-constrained environments or scenarios where data sovereignty is paramount.