TypeSafe AI launches Jev, a non-text 'System One' model claiming 70ms to 500ms responses and free output tokens
- TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev in early access as the first of its System One Models, a class that skips string generation and returns type-safe structured values the company says cannot make type errors.
- TypeSafe claims Jev answers System One queries in 70ms to 500ms against 3 to 329 seconds for frontier LLMs, priced at $0.042 per million input tokens with output tokens free.
- The stack trains with Reinforcement Learning for Calibrated Decisions (RLCD) and a parallel sampler that emits all outputs in a single forward pass instead of one token at a time, which TypeSafe says makes Jev two orders of magnitude faster and cheaper.
- Every Jev output carries calibrated confidence scores, and TypeSafe pitches the model for smart if-statements, map-reduce over large data, real-time apps, and verifying or guarding LLM prompts, reasoning traces, and outputs.
- TypeSafe says it deliberately does not publish performance against public benchmarks and plans only one-off evals at product updates, a choice Hacker News commenters read as a sign the numbers would not look good.
Hacker News opinions
The Doom demo is what sold me, genuinely cool to watch a model play that in real time.
Except when they tell it not to fire and just dodge, it doesn't dodge at all, it walks right up to the pink demon instead of keeping distance. Or am I reading the demo wrong?
Took me way too long to realize Diogo Almeida isn't a joke version of Dario Amodei. Wasn't until the demo videos that I understood the post wasn't satire.
Wild that it doesn't generate text at all. I want to know how this stack compares to Tesla's FSD approach.
Signed up for the beta. My guess is it could replace 40 to 70 percent of the LLM calls in a given pipeline, cutting the API cost on those calls by an order of magnitude.
The line about extraordinary claims requiring extraordinary evidence is exactly the attitude I want from a model release. But the evidence isn't actually there.
What is going on with the outfit changes in the launch video? The whole thing looked AI-generated to me, the voice sounds synthetic.
You could use this for coding if you fed it an AST. If anyone at TypeSafe is reading, please try that.
The CEO actually replied on that: the hard part for coding is state engineering, getting your dependencies into context. He says they haven't tried it yet and want to automate easier tasks first before coding releases.
It never shows how you actually use it, just animations of it working. I want to see the real code behind the demos.
Someone made a dspy fork with a decorator that swaps in TypeSafe where possible on Signatures, that shows a fair bit of hands-on usage. Still rough but it's something.
I'd love this on OpenRouter or AWS Bedrock. Adding a brand new vendor means a whole compliance and purchasing review, but capabilities added to a vendor I already have get adopted instantly. An extra middleman tax is worth it when the savings are one or two orders of magnitude.
All the claims read like marketing so far. RLCD and parallel sampling have nothing backing them up, and 70 to 500ms versus 3 to 329 seconds is apples to oranges unless the LLM baseline is doing comparable work with long chain of thought. If Jev skips generation entirely for a narrow structured task, of course it's faster. Still, I want it to be true.
They do have benchmarks, like the wikipedia page-to-page game. Jev takes the same or fewer hops but roughly 10x less time and 10x less money. Comparing against LLMs doing chain of thought is fair if performance is comparable.
I'm confused by the pricing. LLM output tokens are 5x input, theirs are free?
They aren't doing autoregression, so all outputs are computed in one big forward pass. That's genuinely cheap. It's their output tokens that are free, under the System One column.
They say they deliberately don't publish results against public benchmarks, only one-off evals on product updates. I bet they'd publish if their scores were good.
Why System One? They never explain what a System One task or a System One shaped query even is. Does that mean a fast response, and does it imply a very small model? There's nothing about the model itself.
That page is painful to read. All the typography looks SVG-rendered, semi-transparent gray with a black outline. Never seen that even on the worst vibeslopped sites.