A new model called Jev, released weeks ago by Typesafe AI, does one thing: it accepts text input and outputs floating-point numbers. No sentences, no chat, no essays. Those numbers represent answers to yes/no questions, ranked choices, and scoring tasks. It is extremely fast and extremely cheap precisely because it does nothing else.
Simon Willison published a concise technical breakdown of Jev that explains why this narrow focus matters for production classification pipelines. The model architecture forces a design constraint that turns out to be useful: when you only need a decision, you do not need a language model, you need a decision model. That distinction is what the full write-up is worth reading for.
The category is already expanding. Kev and Nimble are independent follow-on implementations. Nimble is now listed in the Ollama model library and runs locally on consumer hardware. If Jev is the proof of concept, Nimble is the first sign this becomes infrastructure.
[READ ORIGINAL →]