The era of forcing Large Language Models to perform trivial classification tasks is ending. TypeSafe AI and Convai Innovations have released Jev and Laya, two new "decision models" designed to answer bounded questions with probabilities rather than generating prose. Jev, introduced on September 15, saw nearly 13% of Vercel’s paid teams adopt it within 24 hours, marking it as the gateway’s fastest-adopted launch. Laya follows closely as an open-weight alternative under the Apache 2.0 license, allowing developers to run these judgment engines on their own hardware.

The End of the Token-Wasting Classifier

Traditional LLMs are overkill for tasks like routing support tickets or assessing urgency. They generate token-by-token, often producing verbose JSON that applications must parse and discard. Jev and Laya invert this workflow: the application supplies the state and defines the possible answers in advance. The models return choices, scores, or "noul" (yes/no probabilities) directly. This approach eliminates the hallucination risk associated with free-form generation, as the model cannot invent an option outside the fixed set provided by the developer.

Managed Speed Versus Open-Weight Control

Jev operates as a managed API using a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). It supports up to 255 options per choice question and boasts response times between 70 and 500 milliseconds. Conversely, Laya exposes its weights for local inference, clocking 39.5 milliseconds for single questions on a Tesla T4 GPU. While Jev offers convenience and broader option support, Laya provides inspectability and fine-tuning capabilities, particularly useful for organizations requiring strict data governance or specialized domain adaptation.

Accuracy and the Calibration Trap

Speed is irrelevant if the judgment is wrong. Laya’s specialized checkpoint achieved 76.6% accuracy on synthetic benchmarks, but its general English checkpoint scored only 36.2%, falling below the majority-class baseline. This highlights the critical need for calibration; a model can be confidently wrong. Runware staff engineer FΓ©lix Sanz demonstrated that breaking complex judgments into smaller, specific signals can boost accuracy from 62.6% to 95% when combined with simple code logic. Developers must treat these models as probability engines, not truth-tellers, and rigorously test thresholds against real-world consequences.

Key Takeaways

  • Jev and Laya represent a shift toward specialized, non-generative models for agent workflows.
  • Jev is a managed service optimized for speed and broad option handling; Laya is open-weight for local control and fine-tuning.
  • Both models eliminate text-generation hallucinations but require careful calibration to avoid confidently wrong decisions.
  • Integration via Runware’s /v1/systemone endpoint allows developers to swap between managed and self-hosted decision engines seamlessly.

The Bottom Line

Decision models like Jev and Laya are not replacements for LLMs, but they are the missing infrastructure layer for efficient agent routing. By stripping away unnecessary generation, they offer a pragmatic path to lower costs and higher reliability for bounded tasks, provided developers respect the hard limits of their calibration.