Laya

Learning Objectives

• Explain what Laya does and how it relates to TypeSafe's Jev
• Choose between its three checkpoints for a given language and workload
• Read its published results correctly, including which ones come from fine-tuning and which from the base models

What Is Laya?

Laya is an open-weight decision model released on September 18, 2026 by Convai Innovations, a small company in Kasaragod, Kerala, India, led by chief executive Nandakishor. It is a direct response to Jev, the System One model TypeSafe AI launched three days earlier: the same idea, that most AI inside software should return a typed decision rather than prose, delivered as weights you can download and run yourself.

You give Laya a block of state, such as a support ticket, an email or a record, plus a set of typed questions. It answers every question in a single pass of the model, and each answer arrives as a structured value with a probability attached: a choice from options you define, a score on an ordered scale, or the probability that a yes-or-no question is true. Nothing is generated as text, so there is nothing to parse.

Under the hood it is not a chat model at all. Laya is built on an encoder, the kind of bidirectional model long used for classification, with a small decision layer trained on top. The English checkpoint uses ModernBERT-large and totals 421 million parameters; the multilingual one uses mmBERT-base at 322 million. That is why it is fast and small enough to run on one ordinary GPU, and it is also the strongest evidence yet for the most common criticism of Jev, that a model of this kind is a well-packaged classifier rather than a new category.

The code and the weights are released under Apache 2.0, a permissive license with no revenue threshold, no user cap and no regional exclusion, so commercial use is allowed. The GitHub repository reached several thousand stars within its first three days.

The Three Checkpoints

A built-in router detects the script and language of each input and sends it to the right checkpoint. Loading only the checkpoints you need keeps memory down; switching languages with nothing preloaded forces a reload that takes several seconds, so production deployments should preload.

pip install laya

import laya
agent = laya.load("convaiinnovations/laya")          # English checkpoint
result = agent.predict(state, questions)             # all questions, one pass

How Good Is It?

The model card is unusually candid, and the honest reading needs three of its numbers side by side.

On speed, the gap is large and the measurement is direct. The developer reports about 33 milliseconds for a single decision on one Nvidia T4 GPU, and between roughly 100 and 330 decisions per second batched. Independent testers have measured Jev at 236 to 276 milliseconds through its hosted gateway. Those are different setups, a local GPU against a network call, but the difference is too large to be an artefact of that alone.

On accuracy, it wins some tasks and loses one badly. Against published Jev figures, Laya reports 95 percent against 91 percent on a four-label news classification test, and 60 percent against 48 percent on a six-label emotion test. On a banking test with dozens of labels it scores 43 percent against Jev's 87 percent, because each option shares a fixed token budget and dozens of options leave only a few tokens apiece. Raising that budget is documented, but the default loses.

The headline typed-decisions result comes from fine-tuning. Laya scores 76.6 percent against Jev's 72.7 percent on TypeSafe's own four-workflow evaluation. That figure belongs to the third checkpoint, which was trained on that benchmark's own training split. The two general checkpoints score about 36 and 34 percent on the same test, below the 46 percent you would get by always picking the most common answer. So the result shows what a small model can do once fine-tuned on your task, not what Laya does out of the box. Jev also keeps the edge on soft accuracy, meaning how closely its full probability spread matches the reference answers.

Calibration needs work on your side. The models ship over-confident. The developer's best calibration figure is reached only after fitting a temperature adjustment for each question type on in-domain data, and the card tells you to do the same before trusting any threshold.

Pricing

There is no hosted service and nothing to buy. Your cost is the hardware you run it on: about 800 megabytes of weights for the English checkpoint, and CPU inference in the hundreds of milliseconds per decision if you have no GPU.

Strengths

• Open weights under a permissive license — commercial use, fine-tuning and redistribution are all allowed, with no revenue or user limits
• Runs where your data lives — suited to regulated settings where tickets, records or messages cannot be sent to an outside API
• Very fast on modest hardware — tens of milliseconds per decision on a single T4 GPU
• Many questions in one pass — every question about a block of state is answered together
• Multilingual — the multilingual checkpoint is usable in 45 of the 51 languages the developer tested
• Fine-tunable — the typed-decisions result shows a small checkpoint trained on your own examples can match a hosted model, and a fine-tuning notebook is published

Limitations and Considerations

• The general checkpoints are weak on unfamiliar tasks — below a most-common-answer baseline on TypeSafe's business workflows until fine-tuned
• Many options degrade it sharply — at default settings a question with dozens of choices loses badly, so large option sets need a bigger token budget or a two-step choice
• Scores are its weakest question type — ordinal ratings are notably less accurate than choices or yes-or-no questions
• Ships over-confident — probabilities need calibrating on your own data before any automatic threshold is safe
• Self-reported results only — no independent evaluation exists yet, and the comparison with Jev mixes different test conditions
• The English checkpoint fails silently on other scripts — it can be confidently wrong on non-Latin text, so route non-English input to the multilingual checkpoint
• Short context — 512 tokens on the English checkpoint and about 1,000 on the others by default, so long documents need chunking
• Brand new, from a small company — released September 18, 2026, and not mentioned on its maker's own website, whose main product is an offline code-security auditor for healthcare software
• It does not write anything — like Jev, it decides among options you define and cannot summarize, explain or draft

Best Use Cases

When to choose alternatives:

• A task with dozens of possible labels → Jev, or a frontier model with structured output
• No labeled data and no time to fine-tune → Jev through a hosted gateway, which does better zero-shot on business workflows
• Anything that must produce text or a justification → a chat model with structured outputs
• Accuracy matters more than cost or speed → a frontier model

Getting Started

1. Install the package and load the English checkpoint, or the multilingual one if your inputs are not in English
2. Start with a single yes-or-no question on data you already have labels for, and measure accuracy yourself
3. Fit the calibration adjustment on a few hundred of your own examples before setting any confidence threshold
4. If accuracy is short of what you need, fine-tune on your own labeled examples using the published notebook
5. For option sets larger than about 20, raise the token budget or split the choice into two steps

Key Takeaways

• Laya is an open-weight decision model from Convai Innovations, released September 18, 2026 under Apache 2.0, answering typed questions with probabilities instead of text
• It is the open, self-hosted counterpart to TypeSafe's Jev, built on small encoder models of 322 to 421 million parameters
• It is much faster than Jev's hosted service, at about 33 milliseconds per decision on one GPU
• Its best accuracy result comes from fine-tuning on the benchmark's own training data; the general checkpoints trail badly on the same test
• It loses sharply on questions with dozens of options and ships over-confident, so calibrate and test on your own data
• Every comparison with Jev is self-reported and mixes test conditions, so it is a first data point rather than a verdict