AivexaNewsSearch
AI news for builders and product teamsChecked every hour

[AINews] Jev: a “System One Model” that only decides/classifies/routes/scores — >100x faster, >200x cheaper than small frontier LLMs

Collected Sep 30, 2026

TypeSafe launched a model named Jev, which its team describes as a "System One Model" designed only to decide, classify, route and score rather than to generate text or reason. The launch drew enough attention to sit atop Hacker News for a full day, according to Latent Space, which noted that the company had previewed the work at an AI Engineer pre-launch event the previous month.

The claims attached to Jev are speed and cost: more than 100x faster and more than 200x cheaper than small frontier LLMs. The team positions the model as a complement to slower "System Two" LLMs. Users give up strings and chat output in exchange for three properties the team lists: parallel sampling, "no hallucination," and calibration.

Jev was trained through RLCD, which stands for calibrated decisions, a method Latent Space describes as an area Hugging Face's Clementine had highlighted as an important research frontier, contrasting arenas with LLM-as-Judge evaluation.

Why it matters: Developers whose pipelines need routing, classification or scoring steps can consider a dedicated model that the vendor claims is far faster and cheaper than small frontier LLMs, while leaving generation and reasoning to the slower models it is intended to complement. The source does not detail pricing or availability.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

congrats to TypeSafe!