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

You can now run custom decision models locally on your own hardware, bypassing cloud-based inference costs.

Jared Palmer has released Kev, a family of small decision models built on the Qwen3.5 large language model. These models are designed to be trained and run locally, with options for 0.8 billion, 4 billion, and 9 billion parameters. Kev models are inspired by Jev's Architecture Unmasked and offer capabilities for yes/no, multiple-choice, and rating questions within a single request. The API is designed to be compatible with TypeSafe's System One, allowing local server deployment. Kev supports both CUDA and Apple Silicon hardware, with specific performance notes for Mac users. A web playground is available for testing inputs and observing how option order affects outputs. Pretrained weights are provided, along with the code and data for users who wish to train their own models. The project also includes details on previous generations of Kev models based on Qwen3.

The introduction of Kev models, particularly those based on Qwen3.5, signifies a trend towards more accessible and customizable AI decision-making tools. By providing models that can be trained and run on local infrastructure, Kev democratizes access to sophisticated decision modeling, reducing reliance on expensive cloud services. This shift could empower smaller organizations and individual developers to integrate advanced AI capabilities into their applications without significant financial barriers. The open-source nature of Kev and its compatibility with established APIs like TypeSafe's System One further lowers the barrier to entry for developers.

The availability of different model sizes (0.8B, 4B, 9B) allows for flexible deployment based on hardware constraints and performance requirements, from powerful servers to personal Macs. This granular control over model deployment, coupled with the ability to train custom models, suggests a future where AI decision agents are highly specialized and adaptable to specific business needs. The focus on local execution also raises interesting possibilities for data privacy and security, as sensitive information might not need to leave the user's environment.