Laya is a new open-source decision model that ConvAI says runs 7.8 times faster than Jev on the same tests

ConvAI Innovations released Laya on 2026-09-18, an Apache-2.0 decision model that outputs a calibrated probability over a fixed schema instead of generating text. The company positions it as an open-source alternative to TypeSafe's Jev, and the GitHub repository at NandhaKishorM/laya has gained 713 stars in one day. The Show HN thread reached 360 points in about four hours.

What it does

Laya answers three shapes of question: pick one option from a list, score something on a scale, and answer yes or no with a probability. ConvAI calls these choice, score, and noul. The example use cases named on the product page are email spam filtering, phishing detection, jailbreak detection, guardrails around another model, RAG passage relevance, support ticket routing, and invoice processing. Because the output is a probability over the schema you define at call time, the model cannot return anything outside that schema, and ConvAI writes that "hallucination is not possible" by construction.

What the numbers say

ConvAI publishes a benchmark table on the product page comparing Laya with Jev directly. On its own typed-decisions benchmark, ConvAI reports Laya at 0.766 accuracy against Jev at 0.727. On AG News four-way classification it reports 0.950 against Jev's 0.910. On expected calibration error the reported numbers are 0.081 for Laya and 0.246 for Jev, and on median latency the page reports 32.8 milliseconds for Laya against 236 to 276 milliseconds for Jev on the same hardware. The multilingual sweep is described as 45 of 51 languages exceeding three times the random baseline. Every one of those figures is ConvAI's own, and none has been reproduced by an independent party yet.

How it is distributed

Laya is Apache-2.0 licensed, and the model weights, training datasets, and inference code are all in the public repository. Self-hosting is the intended path: ConvAI states there is no per-token cost, against Jev's API price. A hosted endpoint is offered as an alternative.

The founder, Nandakishor Mukkunnoth, published the initial research in March 2025 under the name "Latent Attention Encoder", and the September 2026 release is the productised version of that architecture. The architecture is a bidirectional encoder that runs the whole input in one forward pass, which is what allows the single-pass classification and the sub-35ms latency claim.

Every number on the comparison page came from ConvAI itself, so a team currently paying per token for a closed classifier should test Laya on its own workload before drawing a conclusion.

Source

• ConvAI Innovations: Laya product page
• GitHub: NandhaKishorM/laya
• Discussion: Hacker News thread