LAYA is the open-source entry in the category this site tracks: a local-first decision engine that returns calibrated probabilities instead of generated text observed. It is also — and this matters for anyone searching the name — surrounded by phonetic neighbors that have nothing to do with it. Our tracker saw interest re-ignite recently (a +211% week-over-week jump in the "laya model" query family), with almost no pages dedicated to the actual model — dedicated-page count was just 2 at capture time observed. Labels throughout: vendor claim / observed / analysis.
Search results for anything "laya"-shaped are a collision course. At capture time the front page for the broader query mixed at least four unrelated products observed:
| Result | What it actually is | The one you want? |
|---|---|---|
| github.com/aayushch/laya | The open-source project this page covers — "open-source, local-first" per the repo description observed | ✓ if you want the model |
| laya.aay.sh | The project's own site observed | ✓ |
| layla.ai | An AI trip planner — different product, similar sound observed | ✗ |
| 1ld.ai ("LAYA AI Assistant") | An assistant product using the same name observed | ✗ (unless you wanted an assistant) |
| layla-network.ai | An offline mobile AI assistant observed | ✗ |
This page is about the first row only. If a trip planner or a phone assistant is what you were after, the rest will read like a different language — because it is a different product analysis.
Every cell carries its evidence status. Where a number has one source, we say so; where we can't verify, we don't print it.
| Field | Value | Evidence |
|---|---|---|
| Type | Non-autoregressive decision engine — System 1 shape: options in, calibrated probabilities out | Repo positioning + our ecosystem notes observed |
| License | Apache-2.0, open source | As recorded in our ecosystem notes; verify in-repo for the current text observed |
| Size | ~421M parameters | Per a hands-on review headline on dev.to — single third-party source vendor claim |
| Response profile | ~33ms, described as multilingual | Per a web directory description of the model — not independently measured vendor claim |
| Intended tasks | Text classification and closed-set decisions | Directory description + third-party write-ups observed |
| Runs where | Locally, self-hosted | "Local-first" in the repo's own description observed |
| Context window / pricing | Not stated here — no source we would defend | — analysis |
The defining choice is architectural: LAYA never enters the token-by-token writing regime observed. You supply input text and a closed set of options; it returns a probability for each option, in one pass. A YouTube walkthrough circulating with the project describes the same shape — "open-source AI routing with calibrated probabilities" observed. The consequences line up exactly with the category pattern we explain in What is a System One Model?: no malformed-JSON failure class because no string is written; no invented answer outside your option set because the output space is your option set; and confidence arrives as a number you can threshold on, rather than prose you have to parse analysis.
The trade is symmetric and worth stating plainly: a 421M-parameter decision engine will not write your emails, summarize your documents, or reason through a novel problem. It classifies. If your job is generation, you want an LLM; if your job is millions of small, closed decisions — routing, triage, moderation labels — generation is overhead, and that is the gap LAYA aims at analysis.
Pick the open path when data cannot leave your infrastructure, when per-call cost at volume is the buying criterion, or when you want to audit the whole stack. Expect community-scale maturity: this is a project with a hands-on review trail, not an enterprise SLA analysis. Stay with hosted options when you need breadth of capability in one API — and note that even commercial System One models remain young; our Jev coverage carries the same caution. If you are weighing this against mainstream API models, our pages on GPT Sol 6.1 and the Opus 5.5 vs Fable comparison cover that other side, and the schema generator shows the text-plus-options contract a decision engine of any brand expects analysis.
Open source and self-hosted — our notes record Apache-2.0. Free within license terms; verify current license text in the repo observed.
No — Layla.ai is a trip planner. See the disambiguation table above observed.
A model that returns fast, calibrated decisions over a fixed option set instead of generating text — the category this site tracks. Start with the explainer analysis.
~421M parameters per a third-party hands-on review — small by design, single-source, treat accordingly vendor claim.
This page is descriptive and independent, last reviewed October 5, 2026. Not affiliated with the LAYA project or any product sharing the name. All third-party figures are labeled with their source; performance claims are not independently verified.