A researcher who published work on non-autoregressive decision models in March 2025 has released Laya, an open-source system designed to make instant, structured predictions without generating text. According to the source, the developer spent months building and publishing the initial work—including an arXiv paper (arXiv:2503.23303), model weights on Hugging Face, an open dataset, and a PyPI package.
The frustration behind Laya’s release stems from September 2026, when TypeSafe AI, founded by Diogo Almeida (a co-inventor of ChatGPT at OpenAI), launched Jev using what the researcher describes as the same core concept. The developer claims Jev launched without technical papers, open weights, or training datasets.
Laya operates as a System 1 decision model—designed for fast, reflexive judgments rather than complex reasoning. Instead of using generative large language models for structured classification tasks (like routing customer tickets or detecting spam), Laya outputs calibrated probabilities directly. The system never generates text, making hallucination and JSON malformation impossible by design.
According to the source, Laya runs in 32.8 milliseconds on a single GPU (7.2 ms per question in batches), compared to Jev’s typical 150 ms response time. The system supports three decision primitives: “choice” (selecting from a set of options), “score” (placing input on an ordinal scale), and “noul” (binary true/false questions with calibrated probabilities).
The release includes three specialized model checkpoints: the base English model (ModernBERT-large, 421M parameters, 512-token context), a multilingual version (mmBERT-base, 322M parameters, supporting 100+ languages), and a typed-decisions checkpoint optimized for structured tasks like invoice processing and security alerts. All weights are released under Apache 2.0 open-source licensing with zero API subscription costs.
A critical finding from testing across 51 languages revealed that English-trained models fail catastrophically on non-Latin scripts while maintaining high confidence scores. The Khmer checkpoint achieved 0% accuracy while reporting 95.2% confidence; Hindi achieved 10% accuracy at 94.1% confidence. To address this, Laya includes a Router that inspects Unicode scripts and Latin stopword distributions to automatically select the appropriate model before inference, adding less than 2% overhead.
Key facts
- Laya claims 6-8x speed advantage over TypeSafe Jev (32.8 ms vs 150 ms response time)
- The researcher published initial work on non-autoregressive RL-guided decision models in March 2025
- Laya supports 100+ languages and runs entirely open-source under Apache 2.0 licensing
- English-trained models showed 0% accuracy on Khmer text while reporting 95.2% confidence, revealing the need for automatic model routing
- The system outputs only structured probabilities and numbers, making text generation and hallucination physically impossible