Laya the open source version of Jev
Convai Innovations released Laya, an open-source Apache 2.0 family of non-autoregressive decision models claiming 6-8x faster inference than TypeSafe AI's Jev.
Convai Innovations launched Laya, a family of non-autoregressive 'System 1' decision models that output calibrated probabilities over schemas instead of generating text. It ships three Apache 2.0 checkpoints on Hugging Face: laya (421M, ModernBERT-large), laya-multilingual (322M, mmBERT-base, 100+ languages), and laya-typed-decisions (421M, 0.766 accuracy on security alerts), running in 32.8 ms on a single GPU (7.2 ms batched). The release positions Laya as an open alternative to TypeSafe AI's Jev, a commercial decision API priced at $0.042 per million input tokens with ~150 ms responses. The author cites prior arXiv papers (2503.23303, 2510.01237) and a 51-language MASSIVE benchmark sweep showing English-only encoders fail on non-Latin scripts (Khmer: 0.000 accuracy at 0.952 confidence).
- Three Apache 2.0 checkpoints (421M/322M) bundled in one Hugging Face hub with selective subfolder downloads.
- Primitives choice, score, and noul return calibrated probabilities in a single forward pass with no text generation.
- Claims 32.8 ms per decision on one GPU, 6-8x faster than Jev's ~150 ms responses.
- 51-language MASSIVE sweep shows English BPE vocabularies collapse on non-Latin scripts like Khmer and Armenian.
- Positions use cases include phishing/spam triage, jailbreak detection, ticket routing, and agent observability.
- Tags
- calibrationhugging-facemmbertmodernbertnon-autoregressiveopen-weightsreinforcement-learningtext-classification
- Products
- LayaJevModernBERTmmBERT
- Organizations
- Convai InnovationsTypeSafe AI
From the original article
8 min read at laya.convaiinnovations.com
Everyone in AI right now is talking about a new kind of model: an architecture that is not autoregressive, does not generate text, and gives lightning-fast probability predictions over structured schemas. Seeing the hype online feels both validating and deeply frustrating. I worked on this literally one year back in March 2025.
By nandakishor_ml · laya.convaiinnovations.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.