In plain English
A classifier picks an answer from a fixed list, such as urgent or not urgent. A decision model is a classifier built for software to call, returning each answer with a probability. Tokens are the chunks of text a model reads and writes, and they are what you pay for. Open weights means anyone can download the model and run it on their own hardware.
TypeSafe launched Jev on 15 September as a "System One model": "unstructured state in, typed probabilistic decisions out." It returns structured decisions that software can use directly, each with a probability. Input costs $0.042 per million tokens; output is free.
On 1 October Cloudflare released Clef, which is "fully Jev-API compatible" with open weights under the Apache 2.0 licence. Cloudflare says "the market is getting increasingly saturated with decision models."
The notion of a classifier is important. It sorts: a yes or a no, a one or a zero, without sending the question to an LLM for analysis and spending a lot of tokens. Reach for one when the answers can be listed in advance and you ask the question over and over, such as whether an email is urgent or which team gets a ticket. Use an LLM when you need it to write a summary. My email triage uses both: Jev decides priority, category and action, and a small Llama model writes the summary.
A classifier is one of the dumb little robots: it sorts, and ChatGPT or Claude still do the writing. I expect agent platforms to call them behind the scenes.
Jev has sorted my email for two weeks and I have seen no reports on how well it is doing. If I could get reports back telling me where it's been effective, that would be something different. If you put a classifier in a workflow, log its decisions so you can check them.