A researcher has released Laya, an open-source model designed as an alternative to Jev, a proprietary system built by TypeSafe AI for making fast, structured predictions without generating text.
According to the source, Laya operates as a “System 1” decision model—delivering instant reflex decisions with calibrated probabilities rather than token-by-token text generation. The creator describes spending months developing the foundational concepts, publishing an arXiv paper in March 2025 (arXiv:2503.23303) and a follow-up in September 2025 (arXiv:2510.01237) before TypeSafe AI launched Jev in September 2026 with similar non-autoregressive architecture.
The key difference: Laya is fully open-source under Apache 2.0, while Jev operates as a closed, paid API service at $0.042 per million input tokens. Laya runs in 32.8 milliseconds on a single GPU (7.2 ms per question in batched mode), which the creator claims is 6 to 8 times faster than Jev’s typical 150 ms response time.
Laya handles three decision primitives: “choice” (selecting from a set of options with probability distributions), “score” (placing inputs on ordinal scales), and “noul” (calibrated boolean questions). Because outputs consist only of probabilities and structured data rather than generated text, the system cannot hallucinate or produce malformed JSON.
The model comes in three specialized checkpoints bundled in a single Hugging Face repository. The English version uses ModernBERT-large with 421M parameters, while a multilingual variant (322M parameters) supports over 100 languages. A third checkpoint, “laya-typed-decisions,” targets agent observability and customer service workflows.
One technical challenge the creator highlights involves language support. Testing on 51 languages revealed that English-trained models fail catastrophically on non-Latin scripts—achieving 0% accuracy on Khmer text while reporting 95% confidence. To address this, Laya includes a router that detects the Unicode script of incoming text (requiring under 1 millisecond of overhead) and automatically directs requests to the appropriate model before inference begins.
Laya is available through a PyPI package with zero API subscription costs, according to the source material.
Key facts
- Laya processes decisions in 32.8 ms on single GPU, claimed to be 6-8x faster than Jev
- Fully open-source under Apache 2.0 license with weights available on Hugging Face
- Supports 100+ languages with multilingual checkpoint and automatic script-based routing
- Three decision primitives: choice, score, and noul, outputting only probabilities and structured data
- English model with 421M parameters; multilingual version with 322M parameters