Kev is a family of small decision‑model checkpoints built on the Qwen3.5 base models, released by Jared Palmer on GitHub. According to the repository, the family includes three sizes — 0.8 B, 4 B, and 9 B parameters — each trained with the same data and settings and made available with training code, evaluation data, and serving scripts. The models implement the architecture described in Jev’s “Architecture Unmasked” paper and expose an API that matches TypeSafe’s System One interface, allowing the official Python SDK to be pointed at a local Kev server. Each Kev instance can process a single block of input text (the “state”) and answer multiple questions in parallel. The supported question types are yes/no (noul), multiple‑choice (choice), and rating (score). All questions share the same input but cannot see each other’s outputs, which the source notes encourages calibrated probability outputs rather than hard labels. The server runs on CUDA‑enabled GPUs and on Apple Silicon; the 4 B and 9 B checkpoints fit within a 32 GB Mac when using bfloat16 precision, as documented in the “Serving Performance” section. To run a model locally, users clone the repository, sync dependencies with uv, and launch the server with a command such as KEV_DTYPE=bf16 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009. An example curl request shows how to send a customer‑support ticket and receive structured answers, including probabilities and confidence scores for each question type. The source also provides a TypeSafe SDK snippet demonstrating the same call from Python. Beyond the API, the project ships a Node‑based web playground. After installing dependencies, users can open localhost:3001 to experiment with preset scenarios, edit the input text and questions, and run inference with a single keystroke. The playground includes features that test question order effects, isolated questioning, and delimiter handling. A chess demo lets users treat a board position as the input, legal moves as choice options, and a score question to evaluate the position; games are saved in browser localStorage. Evaluation tables in the README report accuracy and Brier scores on development and test splits. For the 0.8 B model, trained‑source accuracy is 0.829/0.827 and new‑source accuracy is 0.643/0.668, with Brier scores of 0.513/0.473. The 4 B model achieves 0.877/0.870 accuracy on trained sources and 0.794/0.832 on new sources, with Brier scores of 0.316/0.266. The 9 B model reaches 0.876/0.873 accuracy on trained sources and 0.812/0.837 on new sources, with Brier scores of 0.291/0.243. Compared to the hosted Jev model, Kev‑9B trails by about 4.5 points on the new‑source development set. The prior Qwen3‑based Kev generation shows that moving from Qwen3 to Qwen3.5 yields accuracy gains of a few points and lower Brier scores, with the 9 B model improving 7.3 points over its 8 B predecessor on the test set (95 % CI +2.8 to +11.7) and reducing Brier by 0.08. All model weights, tarballs, and SHA‑256 checksums are hosted in the Kev collection and the GitHub release.
