GoKawiilJeeves, a reasoning-based Jev-style classifier, outperforms Jev and Kev-9B benchmarks
Jeeves is a new 9B-parameter Jev-like decision model built on Qwen3.5-9B with LoRA and a pointer head, trained using SFT and CISPO with a block-4 diffusion drafter. It supports yes/no, multiple-choice, and rating questions through a Jev-compatible API, and reports test accuracy of 0.889 versus 0.822 for Kev-9B and 0.857 for Jev, plus a 0.935 score on JevBench's public tiers versus 0.866 for Jev.
GoKawiil's interpretation of the reporting above, not reported fact.
The results suggest that adding a reasoning step before decision-making can improve accuracy on tasks a model wasn't specifically trained for, which could reduce reliance on separate fallback reasoning models in classification pipelines. However, the reported transfer score (0.746) trails Jev's 0.800, indicating the gains may not be uniform across all out-of-domain tasks. The added latency—3.3 seconds median with thinking versus 0.3 seconds without—points to a tradeoff between accuracy and speed that developers would need to weigh.
- Jeeves combines a Jev-like classifier with reasoning and a diffusion drafter, trained via SFT and CISPO on a 9B Qwen3.5 base model.
- It outperforms Kev-9B and Jev on held-out test data and JevBench's public tiers, but underperforms Jev on transfer tasks like MMLU-Pro.
- Reasoning adds latency, taking roughly 3.3 seconds per request versus 0.3 seconds without, though chain length can be truncated to speed it up.
NVIDIA H100 GPU workstation reference guide — This article is deep in the weeds of training and running large reasoning models like Jeeves on H100 hardware, so anyone experimenting with these workflows will want solid reference material on H100-class GPU deployment. A good technical guide helps demystify FP8 kernels and CUDA/Hopper specifics mentioned in the piece.
See NVIDIA H100 GPU workstation reference guide on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-09-29
Published there as: “Jeeves. Reasoning improves Jev-like decision models”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.