TypeSafe opened early access to Jev on 15 September. It is the first of what the company calls System One models: you send a state, as text or JSON, and questions, and it answers with values and probabilities instead of text.
Jev cannot write a sentence, which the company lists as the feature. Possible answers are defined before the call, so the model cannot make a type error and, TypeSafe says, there is no hallucination to catch. Three question types: Choice, one item from a list you supply; Score, a position on a rubric; and Noul, the probability that a statement is true.
Pricing: $0.042 per million input tokens, output free, “too cheap to meter” in the blog’s words. Frontier chat models, in the same comparison, run $0.20 to $10 per million input tokens with output at about five times that, and answer in 3 to 329 seconds against Jev’s 70 to 500 milliseconds. Early access is by waitlist.
What developers did with it in week one
The Register’s first-week roundup: mostly fast jokes with real mechanics under them. A clothing try-on re-renders your photo from a list of clothes, at $0.0011 a decision and about 620 milliseconds a call. Someone added an Urgency column to a spreadsheet; others fed it Doom, Tetris and chess, where it lost to the open-weight GLM 5.3 but cost less to run. The name for running all code on decision calls is JevOps.
Mo Bitar asked what the demos do not answer: whether it is any good. Archer Hume probed the API with 10,000 calls and concluded it reads decisions off internal representations rather than predicting text. Andrej Karpathy wrote on X that Jev revealed latent demand, for a single-token model with low latency and acceptable intelligence, that the race to higher intelligence had skipped.
The local version
NobodyWho published a parody post called “Jev in 25 lines of Python” that is also a working implementation. It loads any GGUF model through llama-cpp-python, reads the logits at the last position for the option tokens, and takes a softmax across those. Feed it an email and three choices and it returns 0.031, 0.084 and 0.885 for legitimate, spam and phishing. No training and no API call; the data never leaves the machine.
The post points at fuller versions: OpenJev, now SemIf, runs in the browser over WebGPU and races two paths on one local model, direct logit readout against writing the probabilities as JSON, token by token. Its accuracy table, on 102 public rows, puts Qwen3 0.6B at 44%, MiniCPM5 2B at 68.6% and Qwen3.5 4B at 81.3%, against 88.3% published for Jev.
Where it gets shaky
Arcturus Labs swept a coin-flip question from 0% to 100%: the Choice probabilities collapsed toward the winner instead of tracking the stated rate. Noul did better but undershot, on a 49% coin it said 40%. The direct readout in the local version is a softmax over the options you showed, a number shaped like a probability rather than calibrated confidence. Calibration is the part TypeSafe says it spent two years building, and the one part of this that is not copyable today.
My read
If you run a GGUF model at home, you already have the primitive this class of model is built on. It is worth an afternoon to wire one into a real decision, routing an alert or triaging a feed, and see whether the numbers are good enough to branch on. The claim to distrust is the confidence figure, not the speed.
Sources: TypeSafe’s announcement, NobodyWho’s post, OpenJev, Arcturus Labs’ coin-flip tests, The Register