Science1 publisher3 min readPublished
TypeSafe's first System One model emits typed probabilities in one parallel pass
TypeSafe's Jev returns typed probability fields in a single parallel pass and, the company says, runs two orders of magnitude faster than comparable LLMs. Its published benchmark scores agreement with two large models.
The Scientist · Science desk
What happened
- TypeSafe AI, after two years in stealth, has put its first System One model, Jev, into early access, describing it as a frontier model built for fast structured decisions that software can use directly.
- Jev gives up string generation and emits all of its output probabilities in parallel instead of generating them token by token; TypeSafe describes it as unstructured state in, typed probabilistic decisions out.
- The company says Jev reaches similar intelligence to existing LLMs on these structured tasks while running two orders of magnitude faster and more efficiently.
- The stack behind it is new on three counts: a model architecture, a parallel sampler, and a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint A model that cannot emit strings cannot take over the jobs whose output is prose, so adopting Jev pushes the engineering work into schema definitions and into deciding what to do with a probability.
- capability If the per-call figure survives outside scrutiny, a team can afford a model call at every node of a workflow for what one frontier call costs today, and put the judgement where the branching happens.
- decision Anyone assessing Jev now has to build a labeled test of their own, because a benchmark keyed to two big models' averaged probabilities caps the measured score at what those two models already believe.
- cost Budgeting a multi-year workflow on a price the vendor itself says it cannot show is unsubsidized puts the buyer's unit economics at the mercy of TypeSafe's funding, though the company expects the price to fall.
Read the workflow evaluation closely: it is the most interesting part of the release and the part that needs the most care. TypeSafe assumes there is a correct compute graph, hands every model the same workflow, and scores each one against reference probabilities taken from the largest and most expensive external models it could reach [12]. In the published comparison those references are the average of two models, Astra and Fable [13]. The company says Jev is "off the charts" and claims it is "owning the Pareto frontier for almost 2 orders of magnitude" on that measure [14].
Scoring against big models' probabilities does dodge a real problem. A fixed harness plus a ground-truth label set invites tuning the harness until the numbers move, and TypeSafe says it neither optimizes for a ground truth classification nor lets the harness change [12]. It swaps in a different problem: agreement with Astra and Fable is not the same thing as correctness. Where both reference models are confidently wrong about a record, a model that reproduces their error scores well and a model that gets the answer right scores badly.
TypeSafe is reasonably direct about how narrow the no-hallucination claim is. Jev gives up string generation and is optimized for structured outputs [6]. The post says a type error would be easy to falsify with a single counter-example, and that it is mathematically impossible [11]. That is a property of the output space. Calibration of the probability inside a field against your data is a different property.
A skeptic can check speed and price quickly. Both come with an asterisk. TypeSafe says its published evals are generally run from its laptops on the West Coast, where the service is currently based [9], and a network round trip flatters whoever sits closest to the endpoint. On price, the company wrote: "We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)" [10].
The side-by-side demo comes with its own disclosure. TypeSafe notes the query is highly simplified with descriptive, human-readable keys, and writes that "The relatively shorter input paints our model in an advantageous light" [17]. In the recorded run against GPT-5.6 Terra, which TypeSafe picked as the closest match in intelligence, the two disagreed on one field, "Churn likelihood level" [18][16].
Founder Diogo Almeida, who writes that he helped build the instruction-following methods behind ChatGPT at OpenAI, opens with the premise: "Models have been superhuman at chat for years, so where is all the automation?" [2][3]. Two orders of magnitude is a factor of about 100, so on TypeSafe's own figure a workflow could make roughly 100 Jev calls for the per-call cost of one comparable LLM call [21]. The decoding argument works without Jev being smarter than anything: autoregressive decoding pays for every token of a structured answer, while emitting the probability fields in parallel pays once [7].
That leaves two checkable claims: the design claim and the cost claim. The parity claim rests on evals TypeSafe designed and ran, scored against reference models TypeSafe selected [12][13]. Anyone with early access can settle it on their own labeled data, and until someone publishes that, I would treat the speed figure as verifiable and the intelligence figure as a hypothesis with a conflict of interest attached.
What to watch
- An outside run of the four published workflows scored against labeled ground truth instead of the Astra and Fable average.
- Whether the per-call price holds after early access, given TypeSafe says it cannot show the pricing is unsubsidized.
- Latency measured from outside the US West Coast, away from where TypeSafe currently hosts the service.