TypeSafe AI said in mid-September that it had released Jev, a model that does not generate text. In the company’s announcement, Jev is described as a "System One model" that takes a state and a predefined set of questions, evaluates them in parallel, and returns typed answers with probabilities. It does not produce strings during the process. Less than a week after launch, several open-source follow-up versions had already appeared.
State in, predefined questions in, probabilistic answers out
Conventional language models generate replies token by token after receiving context. Even when the final output is only a small JSON object, each token still depends on what came before, and the application layer has to parse and validate the result afterward.
Jev is built around a different setup. According to TypeSafe’s technical documentation, both the question format and the answer type are declared in advance. The model supports three basic types:
- Choice: selects one option from a list and returns the selected result, the probability for each option, and a confidence score;
- Score: rates a state against a scoring standard and returns the score, a probability distribution, and a confidence score;
- Noul: answers whether a statement is true and returns a value between 0 and 1.
The documentation says the three question types can be mixed within the same API call and evaluated in parallel. It also says that adding more questions "barely changes response time" and does not cause "context degradation."
Probability is central to the design
The model’s probabilistic output changes how applications can be written. TypeSafe used ticket routing as an example. If the model assigns 91% to the engineering team and 9% to the billing team, automatic routing is a reasonable choice. If the split is 52% versus 48%, the program can escalate the case, add more context, or call a stronger model instead.
Under that design, thresholds and downstream actions are set by software, while the model supplies semantic judgments that a normal if statement cannot derive directly from unstructured text. On its website, TypeSafe describes this as "decisions, not strings" and as typed output that software can act on directly.
The company also notes a limit to that type safety. It can block formatting errors, but it does not prevent judgment errors. The model cannot return an option outside the declared range, yet it can still choose the wrong option within that range.
Speed and pricing claims come from TypeSafe
Figures listed on the TypeSafe website include 0.114 seconds of latency for a System One task, a cost of $0.000081 per task, and $42 per billion input tokens. The company also says input pricing is 238 times lower than Claude Fable 5.1, and that the model is "193.6 times faster" and "444.6 times cheaper" overall.
In a separate announcement, TypeSafe gave a response-time range of 70 milliseconds to 500 milliseconds, compared with 3 seconds to 329 seconds for current frontier models.
Those numbers are based on TypeSafe’s own published tests. The report did not cite any independent third-party replication. The company said Jev is built on a new architecture, a parallel sampler, and a training method called RLCD, short for Reinforcement Learning for Calibrated Decisions. The model is currently in early access, and developers need to join a waitlist.
Open-source projects appeared within a week
The category’s spread has been quick. Based on submission times on Hacker News, open-source alternatives that run on local GPUs, implementations that simulate Jev-like behavior with existing language models, and latency testing projects began appearing from Sept. 16.
Hugging Face also shows models moving in the same direction, including system-one-qwen3.5-4b-scorer, listed on Sept. 16, and cua-s1-forms, listed on Sept. 18 and positioned around computer operation tasks. The latter collected 57 likes within two days of listing, and community-converted ONNX and CoreML versions had already appeared.
The debate around Jev is no longer limited to whether one product works well. It has expanded to a broader question: whether separating judgment from text generation is strong enough to stand as its own model category.