Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 (github.com)

The launch of Kev, a family of small decision models built on the Qwen3.5 architecture, marks a significant development for the AI/ML community. Inspired by Jev's Architecture, Kev offers models with varying sizes (0.8B, 4B, and 9B parameters) that can be trained locally or utilized with pretrained weights. These models allow users to handle complex decision-making tasks by answering yes/no, multiple-choice, and rating questions simultaneously, enhancing the versatility and efficiency of AI applications. Technical details highlight Kev's unique capabilities, such as running on CUDA and Apple Silicon, with specific configurations allowing the 4B and 9B models to fit on systems with as little as 32 GB of memory using bf16 precision. The models are designed to manage stateful input while ensuring each question's independence in processing, thus providing nuanced probability outputs for more accurate decision-making. An accompanying web playground facilitates user experimentation with input variations, deepening understanding of how question order can impact responses. The combination of low-latency performance and the ability to deploy locally opens new avenues for developing decision AI, making Kev an exciting addition to the current landscape of AI/ML tools.