Open-Source Jeff Models Deliver ~30 ms System 1 AI Decision-Making Locally
Developer firelex has released "Jeff," a series of open-weight 0.8B and 2B fine-tuned Qwen and Gemma models engineered for ultra-fast zero-shot decision classification. Operating in a single forward pass without autoregressive text generation, the 0.8B model delivers decision latencies as low as ~28 ms while matching commercial System 1 models like TypeSafe Jev on benchmark tests. This project proves that high-speed, structured decision-making can run completely locally on edge devices such as laptops without relying on cloud APIs or large closed models. It enables developers to integrate real-time classification, routing, and game control directly into applications at software speeds of up to 50 decisions per second. The Jeff 2B model scored 83.1% on a five-benchmark panel, slightly outperforming Jev's published 83.0%, while all synthetic training data was generated locally using open models without cloud services. While highly effective at fast judgement calls, the models are intentionally limited in multi-step reasoning performance compared to larger LLMs.
## BACKGROUND
System 1 decision models are specialized AI architectures that evaluate pre-defined choices and return calibrated probability scores instantly, contrasting with System 2 LLMs that perform chain-of-thought text generation. Commercial implementations like TypeSafe Jev popularised this approach to dramatically reduce API costs and response times for deterministic decision tasks.