# Laya
Laya is an open, multilingual "System 1 decision engine" created by Nandakishor M (Convai Innovations). It answers in about 33 ms and, like [[Jev]], returns probabilities over structured options instead of generating text.
Laya matters for the story more than for the benchmarks. Its author published the approach in March 2025 (arXiv:2503.23303), with open weights on [[HuggingFace]], an open dataset, a PyPI package and a Reddit write-up. When [[TypeSafe AI]] launched Jev in September 2026 without a paper, weights or data, the author pointed out it was the same non-autoregressive decision idea. A few days later, Laya became the #1 trending model on Hugging Face (Clem Delangue shared it).
Commenters also noted that you can get surprisingly far by training something like ModernBERT-large for typed decisions. The open question is generalization: specialized models trained on known problems do well, but Jev's claim is handling new questions and domains with no task-specific training. That's what open alternatives still need to prove.
## References
- [Laya](https://laya.convaiinnovations.com/)
- [arXiv:2503.23303](https://arxiv.org/abs/2503.23303)
- [Clem Delangue on Laya trending on Hugging Face](https://www.linkedin.com/posts/clementdelangue_the-number-one-trending-model-on-hf-is-an-share-7507906261116542977-NcNP)
## Related
- [[Jev]]
- [[System One Models]]
- [[Kev]]
- [[AI Open Weight Models]]