21 Sept 2026
We show a price on 5,092 software listings, and none of them says when it was true
5,092 of our 12,184 software listings show a price, and none records when it was true. We checked 30 at random on the vendors' own pages:…
An open-weight decision model that answers typed questions with probabilities in one forward pass, about 33 ms on a T4 — and whose own model card shows where it fails while staying confident.
Laya is a family of three open-weight "System 1" decision models from Convai Innovations, published under Apache 2.0 on 18 and 19 September 2026, with the Python package's first release at 04:38 UTC on the 18th. It does not generate text. Given a state — an email, a ticket, some JSON — and a set of typed questions, it returns a choice, a score or a probability for each question in a single forward pass, which puts it in the same category as TypeSafe's Jev and Cua's CUA-S1-FORMS on this board. The English checkpoint pairs a ModernBERT-large encoder with a small decision head trained from scratch, 421 million parameters in all with a 512-token budget per question; a multilingual checkpoint uses an mmBERT-base encoder at 322 million parameters and 1,024 tokens; a third is tuned on four typed-decision workflows. Training uses what the authors call RLCD, reinforcement learning against strictly proper scoring rules, intended to make reporting honest probabilities the only way to maximise reward. The figures are all the authors' own, on a Tesla T4: 39.5 ms for one question on the English checkpoint and 32.8 ms on the multilingual one, falling to 7.2 ms a question in a batch of ten. None is stored as a benchmark score, and the headline comparison with Jev mixes the authors' measurements with third-party Jev figures on different prompts and sample sizes, which the card itself says. What earns the release attention is the part most vendors leave out. The card states that the English checkpoint collapses on non-Latin scripts while staying confident — 0.000 accuracy at 0.952 confidence on Khmer — so confidence gating cannot catch it; that the multilingual checkpoint ships uncalibrated and systematically over-confident, and should have its temperature refitted on the buyer's own data before its probabilities are trusted; and it ships a router that picks the checkpoint from the script before the forward pass for exactly that reason. The company says its underlying research dates from March 2025; the date recorded here is when the downloadable checkpoints appeared. There is no hosted API and no price: it runs on the buyer's own hardware.
| Benchmark | Official | Community avg |
|---|---|---|
| No benchmark scores yet. Be the first to add one. | ||
“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).
Sign in to add a benchmark score for this model.
No questions yet. Be the first to ask about Laya.
Everything here links back to the same verified catalogue. Pick your next stop.