What is Jev?
Jev is a decision model. It reads text you supply, answers typed questions about it, and gives you the answer with a probability attached instead of a written reply. TypeSafe AI announced it on 15 September 2026 as the first System One model. You define the questions in your own code, and Jev answers them and stops.
The two things you send are a state, which is the content you want judged, and a set of named questions. What comes back is a typed value under each name. A support ticket goes in, the queue it belongs in comes out, along with a number saying how sure the model is. Nothing in that response is a sentence a person would read.
Is Jev an LLM?
No. An LLM predicts one token at a time and produces text. Jev produces no text at all, and TypeSafe’s jaggedness page for the current version puts it flatly: “jev-1.13 is not trained to generate text.”
Two more differences follow from that. Jev answers in a single pass, which TypeSafe describes as generating “all outputs in a single query” instead of running a generation loop. The output is already typed, so there is no JSON string to parse and no retry when the parse fails. What you give up is everything that has to be written: drafts, summaries and code all go elsewhere.
The architecture is unpublished, so whether a language model sits underneath is a guess. System One models compared with LLMs covers the cost and failure-mode differences, and Jev architecture covers what is known about the internals.
What Jev returns: Choice, Score and Noul
Three question types, plus a fourth in preview for images.
Choice picks one option from a list you define and returns the winner, a probability for every option, and a confidence value between 0 and 1.
Score places the content on an ordered scale you describe in words and returns a probability-weighted position, so the answer can land between two levels.
Noul answers a yes/no question with one number from 0 to 1. There is no separate confidence value, because that number already is the certainty measure.
Jev 1.14 adds a fourth type in preview. Box finds one thing in an image and returns its bounding box. It only works with the opt-in images header, and TypeSafe says it is not ready for production.
Several questions can ride on one request and are answered against the same state in parallel. Choice, Score and Noul works through real output for each shape, and the exact ceilings on options, levels and context are on the Jev model page.
How to use Jev
Three steps, and the whole thing takes a few minutes once you have an account.
First, get a key. Sign in at the TypeSafe console and create one, then put it in TYPESAFE_API_KEY. Both SDKs read that variable when you construct the default client. Getting access and making your first call has the detail, including what access looked like on 20 September 2026.
Second, install an SDK. pip install typesafe-sdk for Python 3.10 or newer, or npm install @typesafe-ai/sdk for Node.js 20 or newer.
Third, make a call. This is TypeSafe’s own minimal Python example:
from typesafe_sdk import Choice, TypeSafeClient
with TypeSafeClient() as client:
response = client.system_one(
state={"document": "I was charged twice. Please fix this ASAP."},
questions={
"category": Choice(
instructions="What is this ticket about?",
criteria={"billing": None, "technical": None, "other": None},
),
},
)
print(response.choices["category"].choice)Check that last accessor against the SDK you installed. TypeSafe’s quickstart page reads answers back as response.answers[...] while the Python SDK’s own page uses response.choices[...], and both pages are TypeSafe’s.
After the first call the hard part is question design. Narrow questions with explicit criteria beat one broad question, which is the argument how to build with System One models makes at length.
What Jev is good at
Repeatable judgments where you know the set of possible answers before you call. Sorting a ticket into a queue, deciding which model or tool a request should go to, filtering retrieved passages, ranking, scoring and yes/no screening all fit.
The economics push the same way. Output tokens are free and the response is already typed, so the pattern of running several small judgments inside one request stops being expensive. That is why support inbox triage, intent and model routing and semantic reranking are the patterns that show up most often in the examples directory.
What Jev cannot do
It cannot write. Anything ending in prose for a human to read needs a different model.
It is unreliable at counting, arithmetic and anything that treats a date as an ordered quantity, all of which TypeSafe documents on its own jaggedness page. Keep that work in your code.
It reads questions literally, so a scoping word or a double negative you did not think twice about changes the answer. And it treats everything in the state as data rather than as a possible attack, so user-supplied text needs a prompt injection screen in front of it. The full list of nine documented weak spots is on the Jev model page.
Who makes Jev
TypeSafe AI, a San Francisco company that came out of stealth on 15 September 2026 with a $40M seed round led by DCVC. It was founded by Diogo Almeida as CEO, Erik Gafni as CTO and Sasha Sheng as COO.
Almeida’s background is the part that shapes the product. TypeSafe’s docs credit him as a co-inventor of RLHF, the method behind InstructGPT and ChatGPT, and the training method behind Jev is pitched as its sibling aimed at calibrated probabilities instead of preferred text. He told TechCrunch on 18 September 2026 that Jev “is trained exclusively on synthetic data.” More on the company is in the TypeSafe AI glossary entry, and on the method in RLCD explained.
Jev compared with LLMs, classifiers and reasoning models
Against an LLM, the question is what the output is for. An LLM returns a string you parse and a log-probability you have to dig for. Jev returns a typed value and a distribution as the primary output. Jev vs Claude and Jev vs GPT run the cost and latency numbers.
With a fine-tuned classifier the question is whether your labels are fixed. A BERT-style encoder trained on your data is cheaper and runs on your own hardware, but its labels are set at training time, while Jev takes them per request. Jev vs BERT covers the trade in full.
Reasoning models are not really competing for the same slot. They spend tokens on intermediate steps and suit problems that need them, where Jev suits the judgments that do not. Whether any of this deserves a new category name is the argument in is Jev just a classifier.
FAQ
Is Jev an LLM?
No. Jev does not generate text, and TypeSafe’s documentation says the current version “is not trained to generate text.” It takes natural-language input like an LLM does, then returns a typed answer and a probability rather than a string. The architecture underneath is unpublished, so the label describes the interface and the training target rather than the internals.
What does Jev stand for?
TypeSafe has not said. Neither the launch post nor the documentation explains where the name came from, and there is no expansion of it anywhere in the public material as of 20 September 2026. The stable version string is jev-1.13.0, and jev-latest resolves to it until about 21 October 2026. jev-preview already points to jev-1.14.0, released on 14 October 2026, which Jev 1.14 covers.
Is Jev free to use?
No, but output costs nothing. Input is $0.042 per million tokens, which TypeSafe also quotes as $42 per billion, and output tokens are free because the model returns typed values rather than generated text. TypeSafe has not published a free tier. The speed and pricing guide works through a real bill.
How do I get access to Jev?
Sign in at the TypeSafe console, create an API key, and put it in TYPESAFE_API_KEY. There is no waitlist: TypeSafe removed it on 20 September 2026 and opened Jev to everyone.
Is Jev open source?
No. Jev is a closed managed API, and TypeSafe has not released the weights, the architecture or the training details. Community reproductions appeared within days of the launch, built on open models such as Qwen and ModernBERT, but none of them reproduces TypeSafe’s training method, because nobody outside the company knows what it is.
Can Jev generate text?
No. Text generation is the ninth item on TypeSafe’s own list of things the current version does badly, and the wording is that the model “is not trained to generate text.” The usual arrangement is Jev making the routing and gating decisions inside your code, with an LLM writing whatever a person will actually read.