PostHog releases Jeeves decision model with reasoning option

PostHog has released Jeeves, a 9B-parameter model for answering yes-or-no, multiple-choice, and rating questions through a Jev-compatible API. Its repository includes model weights, training code, and train, development, and test data for running the model locally on a CUDA GPU. PostHog says allowing the model to reason before deciding raised its score on the company's test set to 0.889, versus 0.822 for Kev-9B and 0.857 for Jev. That reasoning has a latency cost: the repository reports a 3.3-second median per request with it, compared with about 0.3 seconds without it on one H100 GPU. The company also says Jeeves trails Jev on knowledge tests.