Project profile
Laya
Non-autoregressive choice, score, and noul models with an English checkpoint, a multilingual checkpoint, and a router.
NandhaKishorMWebsiteGitHubDemoDocumentation
Last verified: 2026-09-30
Quick facts
| Project | Laya |
|---|---|
| Creator | NandhaKishorM |
| Type | Decision Model |
| Language | Python |
| License | Apache-2.0 |
| Uses Jev for | Separate model behind a System One-shaped server |
| Open source | Yes |
| Status | Active |
What it is
Laya is an Apache-2.0 Python package. The README describes a router that sends text to an English ModernBERT-large checkpoint or a multilingual mmBERT-base checkpoint, and a server that exposes the same POST /v1/systemone path.
Why it matters
It is the open checkpoint in this directory with an explicit multilingual route, plus a local HTTP shape that existing System One clients can target.
Relationship to Jev
| Hosted Jev | Not called. |
|---|---|
| Wire shape | README: laya.serve exposes the router on POST /v1/systemone. |
| Checkpoints | README table: laya is ModernBERT-large, 421M, English, 512 tokens. laya-multilingual is mmBERT-base, 322M, 100+ languages, 1,024 tokens and up to 8,192 with max_len. |
| Question types | choice, score, and noul through Router.predict. |
Architecture
- Router
- English or multilingual checkpoint
- Typed answers
- Optional /v1/systemone server
Key features
- pip install laya. Optional extras cover HTTP, MCP, LangChain, LlamaIndex, CrewAI, and ONNX.
- A TypeScript tree, laya-ts, is published from this repository.
- The README says the multilingual checkpoint is what the router selects for non-English text.
Performance and claims
The README intro says typed decisions over 100+ languages in a single forward pass at 33 ms, and describes training with reinforcement learning against strictly proper scoring rules.
Jev AI Hub confirmed the checkpoint table, the server path, and the GitHub repository record on 2026-09-30. It has not measured latency or accuracy.
Use the checkpoint table for size and language. Use your own labels before you treat a Laya probability as a drop-in for a threshold tuned on hosted Jev.
Limitations
- The README's long-document note says 16 to 18 of 20 requests were correct up to about 4,000 tokens, and that results vary beyond that. It tells you to check long documents on your own data.
- The intro line states 33 ms. That line does not name the hardware.
- This hub has not installed the package or called the server.
Code and repository
https://github.com/NandhaKishorM/laya
GitHub snapshot captured 2026-09-30: 28,825 stars, 2,513 forks, 114 open issues. Last push 2026-09-29. Counts change; this is not a live lookup.
Related Jev concepts
Related guides
Related projects
Sources
- Laya READMENandhaKishorM · accessed 2026-09-30 · github
- NandhaKishorM/laya repository metadataGitHub · accessed 2026-09-30 · third-party