Jev
TypeSafe AI's closed decision model, released 15 September 2026.
- What follows is mostly an inventory of what is not known.That is not an editorial choice. TypeSafe has published a launch post, a docs site and a price, and almost nothing about the model itself.
What kind of model is it
Unknown.source
TypeSafe's launch post says they built “a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)”. That is the whole of it. The words transformer, attention, encoder, decoder and parameters appear nowhere in the post or the docs.source
TechCrunch describes Jev as “a new transformer-based model… that is not a large language model”. That is a reporter's wording, not TypeSafe's, and it is the only architecture claim in print anywhere.source
So: it is not a language model, it does not generate text, and beyond that nobody outside TypeSafe has said what it is.
How it reads the state
About 64,000 tokens per request, of which 32,000 may be the state.source
How it represents them is not published.source
How it turns options into numbers
Not published.source
TypeSafe describes a “parallel sampler” that produces every answer in one query rather than token by token, which tells you the shape of the computation and nothing about the mechanism. Up to 255 options per choice question are supported.source
Compare the Laya page, which answers this question in two sentences, because the weights are downloadable.
How it was trained to be calibrated
RLCD, Reinforcement Learning for Calibrated Decisions.source
The name is published. The method is not: no reward function, no data, no procedure. Laya's authors use the same name for their own training and do publish the details, so the Laya page can describe a concrete algorithm where this page can only report a name.
What it costs to run
- Access
- A paid API, from TypeSafe directly or through Vercel AI Gateway. Signups were paused shortly after launch.source
- Price
- $0.042 per million input tokens. Output is free, because there is no generated text.source
- Speed, stated
- 70 to 500 ms per requestsource
- Speed, measured
- p50 of about 305 ms through the Gateway, network includedmeasured here
- Hardware
- None of yours. It runs on TypeSafe's.
What it cannot do
- Run on your machine.No weights, no local option, and your state leaves your network on every call.
- Be fine-tuned.Whatever it knows is what you get.our inference, from there being no published training path
- Be inspected.You cannot check any claim about it against anything.
- Be relied on for throughput.We hit 1,231 rate-limit retries (HTTP 429) and 14 outright failed decisions across our runs.measured here It can also return 529 when overloaded.source
- Return an invalid option.Worth stating as a genuine strength: it cannot answer off the list. It can still be wrong.
What we measured
Three games, zero training, 10 episodes each.measured here
- Highway:survived all 40 seconds, every episode. No crashes.measured here
- Blackjack:matched basic strategy on 77% of decisions (95% CI 73 to 81).measured here
- Snake:ate 1.8 food per episode, against 17.3 for a 15-line greedy script.measured here
- Reading the situation:its answers moved substantially between opposite situations, which is the test for whether it is reading the state at all rather than answering from habit.measured here
- Caution:it kept its lane on 94% of highway decisions and drove near the minimum allowed speed. In Snake it circled rather than dying.measured here
- So: it does not crash, it does not do anything stupid, and it does not win. Safe and unambitious is a real decision-making style, and on a highway it is the right one.
Sources
Both models are days old and changing. Everything here was checked on 23 September 2026 against the sources listed above.