An open-source project called Kev has released a family of small decision models built on Qwen3.5, offering developers a locally operated system that assigns probabilities to multiple-choice, scoring and verification questions. The project provides pretrained adapters, training code and a server designed to work with the same API shape as TypeSafe’s System One.
Kev is available in 0.8-billion, 4-billion and 9-billion-parameter variants. Its repository recommends the 4B model as a starting point, the 9B version when calibration and accuracy matter more than memory, and the 0.8B option for the smallest footprint. The server binds to the local machine by default and has no authentication, so the maintainers warn users not to expose it publicly without adding access controls.
Rather than generating a free-form answer, Kev receives a state—a block of text or structured data—and questions with possible choices. It returns a probability distribution over those options. That format can support tasks such as routing a customer-support ticket across several departments, where seeing uncertainty across categories can be more useful than receiving a single label. A bundled demonstration also uses legal chess moves as choices and asks the model to score a position.
Each checkpoint combines a rank-16 LoRA adapter and a small pointer head with a Qwen base model. The base weights remain fixed during training. The pointer mechanism scores representations attached to each option, then converts the scores into probabilities. For Qwen3.5, whose architecture includes recurrent Gated DeltaNet layers, the server processes each question separately while reusing the cached state. The repository says this preserves isolation between questions.
The released models were trained for two epochs on 10,000 examples from ten public datasets, supplemented by generated policy and rule-structure examples. The project reports separate development and test results, while cautioning that a comparison with the proprietary Jev model is not controlled because Jev’s training data are unknown. It also notes that probabilities can change when choices are reordered, even though asking separate questions together or individually produced nearly identical results in its fp32 tests.
Hardware behavior differs by model generation. The Qwen3.5 variants can use specialized kernels on CUDA systems, but Apple Silicon currently relies on slower reference implementations for DeltaNet layers. Older Qwen3-based Kev models remain available and are described as the faster Mac option.
Kev’s release is therefore best read as a developer project with published implementation details and self-reported benchmarks, not an independently validated decision service. Its practical contribution is a reproducible route for experimenting with compact probabilistic classifiers locally, including the ability to inspect code, download checksummed weights and train compatible models.
