Overview & result
Nandakishor M shows Laya, built on Jev: Open weights that answer typed questions in one forward pass, 33 ms on a T4.
From the original post
Laya by @nandakishorm: Open weights that answer typed questions in one forward pass, 33 ms on a T4.
How Jev fits in the loop
- Laya is an open, multilingual, non-autoregressive model installed with pip or uv.
- It answers typed questions in a single forward pass; the author gives 33 ms.
- A router picks the right checkpoint for each request.
- Optional extras add an HTTP server, an MCP server and LangChain integration.
How to reproduce
- Review the source documentation and repository at https://github.com/NandhaKishorM/laya.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for research and data analysis.
Limitations
- The README's "honest limits" say the base checkpoints are near chance zero-shot on typed-decisions (0.362 and 0.352 against a 0.318 random baseline).
- The 0.766 figure comes from a checkpoint fine-tuned on that benchmark's own training split.
- The author describes it as a fast base to specialise, not a zero-shot decision engine, and warns against boolean-word labels in choice questions.