Jev vs Laya depends on which published table you open.

Laya is Convai Innovations' open typed-decision stack. Jev is TypeSafe's hosted System One model. Both answer a choice, a score, or a yes/no. Jev's weights are not public. Laya's are Apache 2.0. The numbers below are the authors' tables. This desk did not rerun them, and the Laya repository did not call Jev. The short description of Jev is on what Jev is.

What is Laya?

Laya is three Apache 2.0 checkpoints plus a router that picks among them.

Louis Grenard pointed at it on September 20, 2026. The spec is NandhaKishorM/laya. Install is `pip install laya`. The English checkpoint is ModernBERT-large, 421M parameters, context 512. The multilingual checkpoint is mmBERT-base, 322M, context 1,024. The typed-decisions checkpoint is ModernBERT-large, 421M, context 1,024. Weights are on Hugging Face under convaiinnovations. A local run needs no TypeSafe key. The write-up is Laya's open weights.

How do the published scores compare?

The author's quality table says every Jev cell was copied from a published number, and none of those cells were measured in the Laya repo.

Routed Laya against those published Jev figures: typed-decisions 0.766 versus 0.727 on 2,000 decisions. AG News 0.950 versus 0.910. DAIR Emotion 0.595 versus 0.480. Banking77 is 0.425 for Laya on 77 labels versus a published Jev figure of 0.870 on 72 labels. The denominators do not match. JevBench's row in that story is Laya 70.1 against Jev 75.3. The 75.3 is an earlier composite. On the v1.5 board, Jev 1.13.0 prints at 72.1, behind Cygnet at 73.7 and Winnow-12B Q8 at 73.2. This desk does not have a v1.5 row for Laya. The API board opened October 7, 2026, release v1.6.1 under revision v1.7.12, ranks Jev 1.13.0 third at composite 71.5, with capability 77.1. Opened October 8, 2026, revision v1.7.21, that board ranks 26 hosted systems and Jev 1.13.0 is fourth at the same 71.5. This desk does not have a v1.6.1 row for Laya. Both boards are on the Jev benchmark page and in the JevBench story.

A later 100-item post from the same conversation scores Laya 92 and Jev 100 on the author's own items. The author says it is not a full benchmark. That page is the 100 items. Side-by-side tables, including this pair, are on compare.

What about speed?

The Laya README's own timing is a T4, and the Jev latency it quotes was measured by someone else.

English, one question: 39.5 ms. Multilingual, one question: 32.8 ms. Batched multilingual: 7.2 ms per question at 10 questions. The Jev p50 of 236 to 276 ms in that README is cited from other posts, not from a call in the repo. An Apple Silicon port, laya-mlx, reports 13.42 ms for the 421M checkpoint and 7.39 ms for the 322M checkpoint on an M3 Max, and 63 of 63 validation answers matched upstream Laya. That port does not score hosted Jev. It is laya-mlx.

What has this desk not checked?

We did not install Laya, and we did not send the Laya prompts to Jev.

Base checkpoints in the README score 0.362 and 0.342 on typed-decisions, against a majority-class baseline of 0.461. Khmer on the English checkpoint is reported at 0.000 accuracy and 0.952 confidence. Those rows are the author's. Jev's weights stayed on TypeSafe's side of the launch. A Snowflake package that does call TypeSafe, with a real key, is Jevflake. Signup state for that key is the signup thread. Open-weight projects as a group are on open source.

Is Laya a copy of Jev I can download?

Laya is Convai Innovations' own stack, Apache 2.0, with weights on Hugging Face. It answers choice, score, and noul. Jev's weights are not in that repository.

Did the Laya repo call Jev?

The README says the Jev figures in its quality table were published elsewhere and were never measured in that repo. The author had no TypeSafe API access.

Which JevBench number belongs next to Laya?

The Laya story cites JevBench at Laya 70.1 and Jev 75.3. That 75.3 is an earlier board. Opened October 8, 2026, the API board ranks Jev 1.13.0 fourth at composite 71.5. On October 7 the same board ranked Jev third at that score. This desk has no v1.6.1 row for Laya. The v1.5 page prints Jev 1.13.0 at 72.1, and this desk has no v1.5 score for Laya.