GitHub - jaredpalmer/kev: tiny Jev-like family of decision models built on top of Qwen3.5 you can train and run on your own
Short summary
Kev is a family of small language models (0.8B to 9B parameters) built on Qwen3.5 that make yes/no, multiple-choice, and rating decisions on text inputs. You can run them locally on your own hardware with an API compatible with TypeSafe's System One, plus there's a web playground for testing.
- Three model sizes available with published accuracy metrics: Kev-9B reaches 85.2% accuracy on unseen datasets and trails the proprietary Jev model by only 3.5 points, while Kev-4B and Kev-0.8B offer smaller footprints for different use cases.
- All models run on consumer hardware (CUDA, ROCm, Apple Silicon) with the 4B and 9B versions fitting on a 32GB Mac in bf16 precision, plus you can train your own weights using provided code and evaluation data.
- Probabilities are calibrated by default and the models output confidence scores alongside answers, reducing confident errors to 4% on new data compared to 8.7% without calibration.
Full summary
Kev is a set of open-source decision models built on Qwen3.5 that you can run locally on your own hardware. They come in three sizes—0.8B, 4B, and 9B parameters—and handle three types of questions in a single request: yes/no, multiple-choice, and rating scales. The models work on CUDA, ROCm, and Apple Silicon, with the 4B and 9B versions fitting on a 32GB Mac using bf16 precision. You get back probability distributions for each answer rather than single labels, which matters because a customer service ticket might legitimately involve returns, shipping delays, and billing issues simultaneously. The API matches TypeSafe's System One, so if you're already using their Python SDK you can point it at your local Kev server instead of their hosted version.
The performance numbers show Kev-9B achieves 85.2% accuracy on test data it wasn't trained on, trailing the proprietary Jev model by 3.5 points on development data but beating it on calibration (confident errors at 4.0% versus Jev's 3.7%). All three models got a second training pass in September 2026 on generated examples—policy cases with explicit day counts and examples where deciding evidence was removed. This pushed Kev-9B from 83.7% to 85.2% accuracy on the test set. The models come with calibrated probabilities by default, using a temperature around 2.1–2.4 fitted to in-distribution data, which cuts calibration error from 0.106 to 0.042 without changing any answers. There's an optional setting to append day counts between dates in the input, which boosts deadline-related questions from 80% to 90% accuracy on Kev-9B.
Setting it up means cloning the repo, installing Python 3.12+ with uv, and running a local server that downloads the model weights on first run. A web playground lets you test different inputs and see how option order affects answers, or you can use the TypeSafe SDK to integrate it into existing code. There's also a chess demo where the board is the input and the model rates position strength. The older Qwen3-based versions remain available and run faster on Macs, but development has moved to Qwen3.5. All weights, training code, and evaluation data are public.