Beyond model prose: Jev and the bet on decisions without text

For three years the AI market has measured its progress almost entirely by the ability to produce ever longer, ever more elaborate text. TypeSafe AI announced Jev in a post on the company blog dated 15 September 2026 — the first independent coverage appeared on 16 and 17 September — making it available in early access as the first of a class the company calls «System One Models». The model gives up prose generation altogether and presents itself as a complement to classic LLMs rather than a replacement: Jev takes in an unstructured, text-only state together with a predefined schema, and returns typed values with their associated probabilities. The documentation lists three decision primitives: Choice (a pick among options), Score (a rating) and Noul (a boolean value). The company is aiming squarely at the points in the software chain where what is needed is fast routing, not a discourse.

For version jev-1.13.0, the documentation gives a price of 0.042 dollars per million input text tokens, with output free, and specifies that the system accepts no image, audio or video formats. On its home page TypeSafe AI claims speeds up to 193.6 times higher and costs 444.6 times lower than frontier models, with latencies between 70 and 500 milliseconds; in the launch post, however, the advantage (put at 40 to 200 times) is confined to «System One type» queries at an equal level of intelligence. The non-autoregressive architecture, based on the Reinforcement Learning for Calibrated Decisions (RLCD) method, produces its outputs in parallel rather than token by token. The limits stated in the documentation are 250,000 tokens per second or 1,200 requests per minute, with a 64k token context window per request and a 32k cap on state plus question.

The picture does contain several uncertain elements that are worth spelling out. In the official launch post, TypeSafe AI openly acknowledges that the 0% figure assigned to type errors in its own charts does not come from empirical measurement but from the structural guarantee of the predefined schema, which rules out a malformed output at the root without in itself guaranteeing that the decision is correct. The performance comparisons, moreover, rest on internal evaluations called «workflow evals», run by pushing competing models through TypeSafe AI's own proprietary adapter, with no results on third-party public benchmarks. Even the announcement of 40 million dollars in seed funding led by DCVC comes from outside sources such as TAO Media and finds no direct confirmation in the company's official channels.

Replacing the creativity of prose with the rigour of a typed schema is a pragmatic insight that answers real software architecture needs. What remains to be seen is whether the probability calibration and the promised speeds will hold up under independent benchmarks and in day-to-day use. — Olya