Jeeves is a reasoning Jev-style classifier with a diffusion drafter, trained with SFT and CISPO. It features a 9B Jev-like model (Qwen3.5-9B, Lo RA, pointer head) that thinks before deciding, using a block-4 diffusion drafter. The model outperforms Kev-9B and Jev on out-of-domain test data (0.889 vs 0.822 and 0.857) and on Jev Bench public tiers (0.935 vs 0.866).

It supports yes/no, multiple-choice, and rating questions via a Jev-compatible API. Inference takes ~0.3s without thinking and 3.3s median with thinking on one H100 GPU. Jeeves achieves higher accuracy on hard Jev Bench items (0.865 vs 0.730) and lower uncertainty on unknowable answers (0.055 vs 0.090).

The model runs on CUDA (Hopper for FP8 kernel) and includes full training code and data. Quickstart requires Python 3.12 and a CUDA GPU. Weights can be downloaded and served via Hugging Face or fused checkpoints.

Example requests show structured outputs for department, escalation, and frustration scoring with confidence and probabilities.