Jev-like models are APIs driven by decision models for use in your own code. You feed them input data, such as a customer's shopping transactions, alongside questions like "is the user looking for a return?" or "which item are they referring to?". Each answer arrives as a probability you can plug into your own "if" statements. There is no token-by-token text generation, no conversation history and no context-stuffing, which the source describes as sidestepping the usual problems of coaxing an LLM into a direct answer, including conversational hallucination. It is also fast, answering questions in parallel, and cheap compared with standard LLMs.

Mukkunnoth wrote in a blog post that he had "designed the Jev architecture one year back", and expressed frustration at being overlooked "because [his idea] was built for a vertical use case": SalesRLAgent targeted one specific problem, while TypeSafe turned the concept into a general-purpose hosted product with servers, documentation, developer kits and pricing. He released his own open-source Jev-style model, Laya, roughly the same day as his post. Adding to the confusion, both parties labelled their models "System 1/One", likely a nod to Daniel Kahneman's Thinking, Fast and Slow.

The clone rush has a shared root: most projects, commercial and open-source alike, take a lightweight variant of Alibaba's open-source Qwen and add layers shaping the input, output and decision procedure. Mukkunnoth's Laya instead builds on encoder-only models from the Bert family. The alternatives materialised quickly enough that the source suggests some may have drawn on his published work, though who copied whom remains unresolved.