Introducing Lev (yogthos.net)

TypeSafe has unveiled Lev, a decision engine that builds on the previous release of Jev, a classifier model capable of making real-time decisions. Unlike traditional conversational models, Lev and Jev utilize a fast "System One" thinking approach, producing probabilistic answers quickly—Lev specifically routes queries between a lightweight encoder for rapid responses and a more complex LLM for higher confidence assessments. This architecture allows users to maintain speed while accessing deeper reasoning when needed, addressing shortcomings observed in standard classifiers. Lev's innovation lies in its hybrid model design, where a swift encoder deals with straightforward queries and defers to a local reasoning LLM, like Qwen3.5-4B, when uncertain. Benchmarks indicate that while the encoder alone scores around 61% on challenging adversarial tests, the combined model achieves a 92.4% accuracy with reasonable efficiency. Lev also includes customizable calibration tools to refine confidence outputs based on specific application needs, thereby minimizing overconfidence tendencies common in simpler classifiers. Overall, Lev’s blend of speed, adaptability, and accuracy represents a significant advancement for practitioners in AI/ML, catering to a broader range of decision-making scenarios.