TypeSafe AI launches Jev, a non-text model that delivers calibrated decisions for developers

TypeSafe AI announced this week the release of Jev, a transformer-based model that departs from the large language model paradigm by returning calibrated probability scores rather than textual output. The shift addresses a growing demand among developers for fast, inexpensive AI that can be trusted in automated workflows, a need that the company says has been “thrilling” the developer community.

The creator of Jev, Diogo Almeida, spent years at OpenAI where he helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF). Despite the breakthrough, Almeida grew dissatisfied with the focus on optimizing human language, describing it as “lightning in a bottle… not useful” for automation. Two years ago he left OpenAI to found TypeSafe AI, aiming to build a model that speaks the language of computers instead of humans.

Jev’s design eliminates text generation entirely. Instead, it produces probability distributions, what TypeSafe calls “calibrated decisions”, that developers can interpret directly. Because output tokens carry no cost and input tokens are billed by the billion rather than the million, the model runs at a fraction of the price of typical LLM APIs. The architecture also sidesteps hallucinations, since the model cannot fabricate text it has never been asked to produce.

Early adopters have reported dramatic performance gains. Pranit Sharma, a software engineer at Vercel, replaced OpenAI’s ChatGPT Luna 5.6 classifier with Jev and observed inference speeds five to eighteen times faster while achieving higher accuracy. The Vercel team used the model to vet command-line inputs for safety, noting that Jev’s speed enabled real-time decision making that was previously impractical.

Another test by Nikhil Mudholkar, CTO of Bryo AI, compared Jev to Google’s Gemini for sorting business emails. Gemini edged out Jev in raw accuracy, but its cost was ten to twenty times higher. Mudholkar highlighted Jev’s confidence scores, calling the model “the only one that hands back a real probability which makes it ideal for automating workflows.” The ability to gauge certainty directly is seen as a key advantage for building reliable pipelines.

Beyond replacing LLMs in specific tasks, Jev can act as a safeguard for larger language models. Almeida envisions developers deploying Jev to monitor LLM agents and detect jailbreak attempts, a use case that would be prohibitively expensive with another LLM. Armin Ronacher, CTO of open-source model-harnessing platform Earendil, agrees that Jev lets users “delegate the hallucination problem” by setting probability thresholds before acting on model output.

The model’s name honors 19th-century economist William Stanley Jevons, whose paradox describes how falling costs drive greater consumption. Almeida believes that cheaper intelligence will proliferate across software much like the early internet, rather than concentrating in a few “mega apps.” He describes the vision as “smart software all over the place in a way that’s emergent and distributed.”

While TypeSafe has kept Jev’s internal architecture confidential, observers suspect it builds on an open-weight LLM foundation. The company describes Jev as a “System One” model focused on intuition over reasoning, trained exclusively on synthetic data through a process Almeida dubs “reinforcement learning from calibrated decisions.” He claims that owning the data pipeline has been “one of the best bets” the company has made.

Looking ahead, TypeSafe plans to extend the approach to additional modalities and expects competitors to follow suit now that the utility is evident. Ronacher predicts a wave of similar models, noting that the current cheapness of LLMs has delayed creative alternatives. Almeida, however, distances TypeSafe from the hype-driven “frontier labs” of the AI sector, emphasizing that the goal is practical intelligence rather than speculative grandiosity.