What Jev actually returns — System One in plain language
Jev is not a smaller LLM with stricter JSON mode. TypeSafe defines a System One model as parallel, typed, calibrated decisions with no string generation at all.
A useful way to read Jev is to stop comparing it to ChatGPT as a product and start comparing it to an if statement that can see messy state.
TypeSafe’s launch post draws a hard line between existing LLMs and what it calls System One models. LLMs are trained to emit strings. Even when you ask for JSON, the sampler still walks token by token, and the surrounding code has to parse, validate, and retry. Jev’s contract is the reverse: the caller defines the legal answers first. The model is not allowed to invent a fourth option, a speech, or a malformed object.
The public primitives are small:
| Primitive | What you ask | What you get back |
|---|---|---|
| Choice | Pick among named options (cardinality up to 255) | A distribution over those options |
| Score | A numeric judgment on a scale you define | A value plus confidence |
| Noul | A yes/no (the docs’ name for a boolean question) | A probability |
Several questions can ride in one request. TypeSafe says sampling is parallel, not autoregressive, which is the mechanical reason it claims 70–500ms end-to-end instead of multi-second chat completions. Input is priced at $0.042 per million tokens; output is free, described as “too cheap to meter” under the new architecture.
The training story is a method they call RLCD — Reinforcement Learning for Calibrated Decisions. Where RLHF optimizes for what a rater prefers to read, RLCD is supposed to make the probability on the label match how often that label is actually right. Calibration is the product feature: software can set a threshold, auto-act when the model is sure, and escalate when it is not.
The blog also argues that schema match is not an eval score. If the output type is defined in advance, a type error is a bug in the system, not a slightly wrong sentence. Hallucinated tool names, extra keys, and “sure, here is a refund” when refund was not an option are the failure modes this interface is trying to make impossible.
What Jev gives up is equally load-bearing. It cannot write an email, a patch, or a chain of thought for a human to read. Almeida’s follow-up in the launch thread is blunt: the speed and cost gains are not free, because the model cannot generate text. Builders who need language still call an LLM. The bet is that a large fraction of “AI in production” was never a writing problem — it was classification, routing, scoring, and “should we do this now?”
The name System One is borrowed from Kahneman’s fast/slow split. TypeSafe is trying to reclaim “fast” from “sloppy”: a reflex layer for software, with the slow model reserved for work that actually needs a paragraph.