PostHog releases Jeeves, an open 9B decision model that reasons before it decides
PostHog published Jeeves on 29 September 2026, with weights on Hugging Face and the full training code and data. The model is Qwen3.5-9B with a rank-16 LoRA and a pointer head, trained first with supervised fine-tuning and then with CISPO reinforcement learning so that it writes a reasoning chain before a trained head scores the options. A single temperature fitted on dev turns the scores into probabilities. Author: Nicholas P. Waltz; the README credits Kev as the inspiration.
It serves Noul, Choice and Score over /v1/systemone, ships jeeves_sdk as a drop-in for TypeSafe’s Python SDK, and runs on CUDA or Apple Silicon. A block-4 diffusion drafter speeds up the chains (176 tokens/s for one question on an H100, per the README). The repo passed about 400 stars in four days.
The numbers (UnverifiedClaim). On JevBench’s 231 public items, PostHog reports 0.935 against Jev’s 0.866 (hard tier 0.865 vs 0.730) and ECE 0.037 vs 0.049. On its own held-out test split, 0.889 vs Jev 0.857 and Kev-9B 0.822, with the Jev and Kev columns taken from Kev’s tables on different items. It trails Jev on MMLU (0.793 vs 0.900). Thinking is on by default: 3.3 s median and 17.1 s p90 on one H100; with thinking off, about 0.3 s and 0.775 on dev.
How we list it. Catalog as a lora-adapter, 6.1 maturity / 6.8 capability / 14 adoption. The GitHub repo says MIT and the Hugging Face weights say Apache-2.0; we flag the mismatch on the page.