Emerging Ecosystem for Fast Decision Models: Jeeves, Jeff, and Jevstiller Released alongside OpenAI's Decisions API

PostHog has released Jeeves, a reasoning-based classifier built on Qwen3.5-9B. Unlike standard Jev-like models, Jeeves uses a reasoning chain and a diffusion drafter to improve performance on out-of-domain tasks. It aims to provide better decision accuracy by performing reasoning before reaching a final classification.

For high-speed, local implementation, an independent project titled Jeff has been released. Based on fine-tuned versions of Qwen3.5 and Gemma 4, Jeff is designed to fit into existing code with the same request format as TypeSafe's Jev. The 0.8B parameter model can achieve decision speeds of approximately 22 ms on an RTX PRO 6000 and 28 ms on an Apple M4 Max. While optimized for speed and zero-shot classification, the developers noted that its reasoning capabilities do not match the larger Jev model.

To bridge the gap between large models and local deployment, Jevstiller has been introduced. This tool acts as a proxy that distills the performance of the Jev model into smaller, local models. It includes a mechanism to ensure that the local model's answers agree with Jev within a user-defined "disagreement budget" using an exact finite-sample bound. In benchmarks using the Banking77 dataset, Jevstiller achieved 71.9% coverage with 99.50% agreement with the teacher model on a CPU.