Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
Kev 1.0 ships four local decision-model checkpoints, from 0.8B to 27B, with a Jev-compatible API.
The models answer yes/no, multiple-choice, and rating questions over the same input text while keeping questions isolated from one another. Kev-4B is the recommended starting point; Kev-27B is the strongest release and is reported within one point of Jev on the listed new-source accuracy comparison, though the README says that is not a controlled architecture test. The repo includes local serving, Modal deployment, browser testing, and fine-tuning paths, with calibrated probabilities enabled by default. It also spells out limits: smaller models validate only to 8,192-token contexts, option order can affect answers, and Kev-27B needs an 80 GB GPU.
HN · GitHub AI's note
The models answer yes/no, multiple-choice, and rating questions over the same input text while keeping questions isolated from one another. Kev-4B is the recommended starting point; Kev-27B is the strongest release and is reported within one point of Jev on the listed new-source accuracy comparison, though the README says that is not a controlled architecture test. The repo includes local serving, Modal deployment, browser testing, and fine-tuning paths, with calibrated probabilities enabled by default. It also spells out limits: smaller models validate only to 8,192-token contexts, option order can affect answers, and Kev-27B needs an 80 GB GPU.
HN · GitHub AI's note
score 5
Discussions
- hn · 463 points · 211 comments