The newest family of decision tools based on the Qwen3 architecture provides a suite of self-hostable models designed for local deployment. These versions, including 0.6B, 4B, and 8B parameter sizes, employ a LoRA and small pointer head technique to mimic Jev's functionality under an Apache 2.0 license. Training the 4B version takes 40 minutes on one H100 GPU, while the 8B version requires 83 minutes.

Kev-8B recorded a 79.6% score on data it was not trained on, trailing Jev's 85.7% on the same set. However, the 4B model exceeded Jev's performance on specific logic recognition tasks, and the open source release allows for local execution without proprietary constraints.

Hardware benchmarks show Kev-4B processing five questions in 300 ms on a 32 GB Mac in bf16 precision or 40 ms on an H100. Integration is handled via a drop-in TypeSafe System One API that works with a single base URL change in the SDK. Repeated documents are served 2 to 2.5 times faster through the use of a KV cache.