Jev introduces a new shape of LLM—System One, aka Decision Models
Short summary
TypeSafe AI released Jev, a model that takes text input but outputs confidence scores and probability distributions instead of generated text. It's designed for classification tasks like spam detection and ranking, costs $0.042 per million input tokens with free output, and has already spawned open-source recreations and creative experiments.
- Jev returns structured probabilistic outputs (yes/no scores, rankings, category choices) rather than text, making it fast and cheap—you only pay for input tokens, not output.
- The model trades interpretability for simplicity: you get a floating-point score but no explanation of which signals drove the decision, raising serious concerns about hidden bias in high-stakes use cases like hiring.
- The community has already built unconventional projects with it (a chat model, a game player, a left-pad implementation) and created open-weight alternatives like Kev within days of release.
Full summary
TypeSafe AI released Jev, a new kind of LLM that breaks the text-in-text-out mold. Instead of generating words, it takes unstructured text or semi-structured data and returns floating point numbers—confidence scores for yes/no questions, probability distributions across multiple choice options, or numeric ratings along a scale you define. The model is remarkably cheap at $0.042 per million input tokens (cheaper than GPT-5 Nano) and charges nothing for output, making it practical for high-volume tasks. It's also fast because questions run in parallel, so asking fifty questions takes roughly the same time as asking one. This makes it well-suited for classification work: spam detection, content labeling, prioritization, and search reranking where you'd use something like BM25 to fetch candidates, then have Jev score them all at once.
The tradeoff is a significant step backward on interpretability. Regular LLMs are black boxes, but at least they can attempt to explain their reasoning. Jev gives you nothing but a number. If it flags something as spam or rates a job applicant, you have no window into which signals drove that decision. This opacity becomes genuinely dangerous when bias is involved—the article notes an informal test where Jev rated Cupertino as a "good city" and East Palo Alto as a bad one, raising obvious red flags about what's baked into the model. The cheapness cuts both ways here: while it makes experimentation affordable enough that you can run thousands of test cases for a few cents, it also lowers the barrier for someone to deploy Jev in high-stakes decisions without proper scrutiny.
The community has already gotten creative with unconventional uses. Someone built a terrible chat interface by asking Jev "which symbol comes next?" iteratively. Another person implemented the left-pad function using a choice query. People are playing 2048 with it. More seriously, open-weight alternatives are already appearing—a project called Kev built decision models from Qwen using 0.8B, 4B, and 9B parameters, and a benchmark called JevBench has cropped up to compare these models. All this activity happened in less than a week, suggesting this format might actually stick around.