TypeSafe launches Jev, a decision-focused model trained with RLCD, claiming orders-of-magnitude speed and cost gains over small frontier LLMs

TypeSafe, a startup co-founded by Diogo Almeida, launched Jev, a model trained with a method called RLCD (reinforcement learning for calibrated decisions) and optimized for decisions, classification, routing, and scoring rather than free-form text generation. The company claims Jev is 20–200x faster and 40–400x cheaper than small frontier LLMs, with output tokens described as free.

Jev is positioned as a complement to slower, general-purpose “System Two” LLMs, offering parallel sampling, no hallucinations, and calibration for structured decision tasks. The team frames it as a new frontier model category tailored for production use cases where autoregressive generation is unnecessary overhead.

Community reactions highlighted Jev’s likely fit as a structured classifier or judge in production systems, drawing comparisons to DSPy-style signatures and typed prediction abstractions. Engineers noted that Jev is not a general language model and requires predefined output formats, making it closer to a constrained or diffusion-like decision model than a GPT replacement.

The announcement received high engagement, topping Hacker News for a day alongside other AI news, and included public resources such as a blog post, evaluation results, and documentation.