HunterAlphaHub
OpenRouter model reference Facts from the public catalogue, dated and labelled
Jev Laya Verified
2026-09-22

Resources · Laya

Laya resources

Everything a reader needs to check this topic themselves: the repository and the three sets of weights, the package we installed to run our own probe, two demos that answer without an install, and the handful of write-ups we went back to. Grouped by what they are for, with a line on each.

Every link opened 2026-09-24 15 resources What we measured with two of them

Official

  • laya (repository)

    Nandakishor M · Convai Innovations

    The model, the three checkpoints, the router, the training code and the Jev-compatible server.

    Our note. Start here. The benchmark table is the publisher's own and says which of its columns came from Jev.

  • Documentation

    Convai Innovations

    Install, the three question types, the checkpoints and the server, on GitHub Pages.

  • laya on PyPI

    Convai Innovations

    The package we installed to run our own probe: `pip install laya`.

    Our note. The 0.3.x line is current. Installing it pulled about 850 MB of weights on first load in our run.

Weights

  • convaiinnovations/laya (weights)

    Convai Innovations

    The English checkpoint: ModernBERT-large, 421M parameters, Apache-2.0, 512-token context.

    Our note. The model card's limits section is unusually direct about overconfidence, and it is right.

  • convaiinnovations/laya-multilingual

    Convai Innovations

    The 322M multilingual checkpoint: 100+ languages, 1,024-token context, the fastest of the three.

  • convaiinnovations/laya-typed-decisions

    Convai Innovations

    The checkpoint the published benchmark is about: four typed-decision workflows, 0.766 accuracy on the publisher's 2,000-question set.

Try it without installing anything

  • Hosted demo

    Convai Innovations on Hugging Face Spaces

    Type a state and a question, get probabilities back, without installing anything.

    Our note. Opened 2026-09-24 and it answered. The quickest way to see the shape of the output before deciding to download weights.

  • Colab notebook

    Convai Innovations

    The quickstart as a notebook, linked from the repository's own badges.

  • laya-playground (live)

    wdobry

    The games and the question editor from the playground repository, with no model behind them.

    Our note. It replays recorded runs and says so on the page. Clone the repository if you want the real model.

Worth reading

  • Builds: what people shipped in week one

    GitHub

    Who forked the model, which is how we found most of the ports in our own builds column.

  • Local Laya vs Hosted Jev

    James Cruce · astgl

    A first-person deployment on a Mac Studio: the gateway design, the fail-closed rules, and the profile that failed acceptance and was switched off.

    Our note. The most useful third-party write-up on this model, and the only one that reports a rule its author disabled after testing.

  • Laya on Mac M4 Core ML

    Hacker News

    The thread where the local-memory question was worked out in public, 174 points.

  • I built non-autoregressive decision models with RL

    Hacker News

    The author's own submission of the year-old work, 1,349 points and 316 comments, most of them about credit rather than about the model.

    Our note. Read the comments as well as the post: the disagreement is the informative part.

  • awesome-jev

    yibie

    A curated list that tracks the open clones alongside Jev itself, and the source of several rows in our own columns.

    Our note. A list of links, not a review. We read it for leads.

  • The OpenRouter directory

    OpenRouter

    Where the hosted models on this site are listed. Laya is not on it — it runs on your machine.

    Our note. Included because a reader looking for Laya in the catalogue will not find it, and that absence is the point of the model.

What this page does not do

Recommend a model, or rank these. A resource index that sorts by quality is an opinion column wearing a list's clothes; what we can vouch for is that each link answered when we opened it, and what the thing on the other end is. Two of the rows carry a note because we actually used them while writing this topic, and the rest say what they are and stop.

Send us a link

A project built with Jev, a post, a video, a correction, a tip. We open the link, check it says what you said it says, and write the entry ourselves.

Required a link, and an email to reply to. Optional everything else.

Add context — all optional

One or two sentences about what it does, in your words. We write the entry ourselves.

Cost, latency, a benchmark — anything you measured. We attribute these to you.

We store what you type and email it to ourselves. No IP address, no user agent, no referrer — the same rule as the mailing list.

What happens to what you send →