Laya (NandhaKishorM/laya)

Laya is an open decision engine published on GitHub by NandhaKishorM on 18 September 2026. It reads a piece of text and a set of typed questions and answers with a label, a score and yes or no probabilities, from one forward pass of a 421M-parameter encoder, generating no text at all. Eleven days on, GitHub's REST API counted 27,682 stars and 2,413 forks, the highest of any AI repository created that month. Stars are attention, not quality.

We read the project's README and three model cards, four editorial write-ups, three news reports from 19 to 20 September 2026, its Hacker News launch thread and three reports on its own tracker, and the GitHub and PyPI APIs, all on 29 September 2026. Nothing was installed, prompted or tested.

Consensus

The eighteen sources agree on three things. Laya is quick: one question in tens of milliseconds on a single T4-class GPU. Laya is open: three checkpoints under Apache-2.0, no per-token bill. And as shipped it is not a decision engine anyone can use untrained. The project says so first, in its README and both base model cards.

That is an odd shape for a release of 27,682 stars: the claim it went viral on is one its own documentation qualifies at length. Confidence is moderate. The documentation is dated and detailed, but two write-ups work from the project's own figures, the one independent test we found runs 78 cases, and no source describes a deployment older than a few weeks.

Recurring strengths

Speed, measured and repeated. Ten of the eighteen sources state the latency figure, and the README attaches the test: 32.8 ms for one question on the multilingual checkpoint, 39.5 ms on the English one, on a Tesla T4, falling to 7.2 ms and 15.9 ms per question at ten in one pass. The independent test that put Laya last on accuracy still timed it first, at 30 ms a case against about 302 ms for Jev.

An open licence and no meter. Seven sources name it: Apache-2.0, three checkpoints, weights on Hugging Face, a package on PyPI. For a team whose state cannot leave its own network, that is the whole argument.

Documentation that names its own failures. Five write-ups single out the README's limits section as the reason to trust the rest of it.

Calibration, once somebody refits the temperatures. Six sources report it. Both checkpoints ship over-confident, and refitting one temperature per question type moves the English checkpoint's mean expected calibration error from 0.466 to 0.081 and the multilingual one's from 0.314 to 0.106.

The narrow binary decisions. Spam filtering at 0.993 and phishing detection at 0.980 in the review we read, with a ten-way routing task at 0.522.

Recurring complaints

Zero-shot, the weights you download are near chance. Nine sources state it and the project's own table is the plainest. On 2,000 typed decisions the English base scores 0.362 and the multilingual base 0.352, against 0.318 for a random guess and 0.461 for always answering with the plurality label. The 0.766 in the headlines belongs to a third checkpoint, fine-tuned on that benchmark's own training split.

Both checkpoints ship over-confident. Six sources raise it; the multilingual one ships with no fitted temperatures at all. A hands-on write-up that ran the model on plain CPU reported seconds per ticket there rather than milliseconds, and one misclassified ticket in four.

Choice questions are the weak primitive. Four sources and the tracker line up. Past about twenty options accuracy falls away, because every option shares a fixed token budget: on the 77-label Banking77 set that is three or four tokens a label and 0.425 against Jev's 0.870. Order matters too: one tracker report found reversing the options and their descriptions changed the selection in 69 of 100 catalog cases on the multilingual checkpoint, 65 on the English one and 76 on the typed one.

The multilingual path is where the seams show. Four sources and one tracker report. The English checkpoint collapses outside English and stays confident while it does: over a 51-language sweep it averages 0.227 with mean calibration error 0.733, only 23 of the 51 clear three times random, and Khmer scores 0.000 at 0.952 confidence. Routed, 45 of the 51 clear it. The router has its own report: its default holds one checkpoint, reloading on every switch, and Spanish and Italian route to the English checkpoint.

Where reviewers split

Speed against accuracy, in two rooms. The project's table puts the fine-tuned checkpoint at 0.766 across 2,000 decisions, above Jev's published 0.727 and above the 0.735 teacher ceiling. The independent 78-case run on a Mac mini put Laya last of four open models at 0.590, behind Jev's hosted API at 0.974, a one-bit 27B at 0.885 and GLiNER2 at 0.795. Both are honest and answer different questions: one suite is the benchmark the checkpoint trained on. Speed held in both.

Whether 7.8 times means anything. The README sets its 32.8 ms against Jev's third-party 236 to 276 ms and calls it 7.8 times faster, saying plainly that it never measured Jev. A comparison piece we read notes that Jev's 70 to 500 ms is an end-to-end figure over the network while Laya's is one local forward pass.

Whether the idea was first. The news report we read describes the author's claim that he published a non-autoregressive, reinforcement-trained decision framework in a March 2025 preprint, a year before TypeSafe launched Jev without papers, open weights or datasets. Commenters in the 318-comment launch thread answered differently: that the concept has academic precursors, that GLiNER is a closer published match, and that the funded lab's contribution was a usable product. Nothing read here shows Jev reusing any code, and a priority claim is not a finding of copying.

Whether the pace is momentum or churn. PyPI lists 29 releases in eleven days, nine uploaded on 24 September and 109,457 downloads in the previous month, and the GitHub API counted 115 open pull requests against 58 open issues, with no push since 27 September. The README's release notes for 0.3.21 describe finished work. Both are true: a repository this fast leaves a busy queue and a long changelog.

Who it suits

You have a few thousand labelled decisions of your own and one T4-class GPU. The project's notebook reproduces the fine-tuned checkpoint in four to five hours on two free T4s, and every source reporting 0.766 also reports what the base checkpoints score without that work.

You need the decision to stay on your own hardware. Three checkpoints under Apache-2.0, one modest GPU, no per-token bill and no round trip.

Who should pass

You want a decision engine that works without training. On the project's own benchmark the base checkpoints score 0.362 and 0.352, against 0.318 for a guess.

Your label set runs past about twenty options, or the order of your options will not hold still. Banking77 is 0.425 against Jev's 0.870 at the default budget, and one tracker report measured the answer changing with the option order.

You will trust the confidence number as it ships. Both checkpoints are over-confident as released, and the multilingual one has no fitted temperatures to trust.

You need long, noisy documents read reliably. A tracker report found zero-shot decisions on long multilingual posts landing near the majority-class baseline.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it.

Sources

• GitHub README, official, read 29 September 2026.
• Hugging Face, the laya model card, official, read 29 September 2026.
• Hugging Face, the laya-multilingual model card, official, read 29 September 2026.
• Hugging Face, the laya-typed-decisions model card, official, read 29 September 2026.
• GitHub REST API, repository figures, usage signal, read 29 September 2026.
• PyPI, laya, usage signal, read 29 September 2026.
• eesel AI, Laya AI review, editorial, 22 September 2026.
• Prolyz, Jev vs Laya, editorial, 22 September 2026.
• Robot World, Laya, the open-source System 1 decision model, editorial, 21 September 2026.
• Ap[e]Chat Blog, Three Jev clones and a 27B, editorial, 20 September 2026.
• AI Weekly, Convai ships Laya, news, 19 September 2026.
• Tech AI Wire, an open-weights answer to Jev, news, 19 September 2026.
• Gadget Pilipinas, the System One model, answered by Laya, news, 20 September 2026.
• Hacker News, the launch thread, forum, 19 September 2026.
• GitHub issue 172, the router's default, forum, 22 September 2026.
• GitHub issue 171, catalog selection, forum, 22 September 2026.
• GitHub issue 99, zero-shot on long posts, forum, 21 September 2026.
• dev.to, replace your LLM-as-a-judge, other, 27 September 2026.

On 29 September 2026 we read the project's README and three model cards, four editorial write-ups, three news reports, a launch thread with 318 comments and three reports on the project's own issue tracker, and the GitHub and PyPI APIs, by web search and plain page fetch, and synthesised them with AI; nothing was installed, prompted or tested.