Jev: The Frontier Model That Refuses to Write

A developer built Jev, a frontier model that never generates text and instead scores a supplied set of options in a single parallel pass, replacing roughly 8.5 seconds of LLM prose with about 0.1 seconds of arithmetic. The project targets the routing, triage, classification and gating workloads that make up most enterprise agent traffic, and the writeup covers the demo, capability trade-offs, and where a decision-only model fits in model routing and tool-call guardrails.

Jev is a frontier model that never generates text. You hand it the options; it scores them in a single parallel pass — roughly 8.5 seconds of LLM prose replaced by about 0.1 seconds of arithmetic. That trade matters for the cost architecture of every agent you run: most enterprise traffic is routing, triage, classification and gating, not essay-writing. The full essay walks through the demo, the honest capability trade-offs, and where a decision-only model fits in model routing and tool-call guardrails. Read the full essay on my site: Jev: The Frontier Model That Refuses to Write https://www.vincirufus.com/en/posts/jev-model-that-refuses-to-write/