Badger Cheese :
Worth checking this against the model card before repeating it.
1. Not Chinese. Laya is from Convai Innovations, a small company in Kasaragod, Kerala, India. Founder is Nandakishor Mukkunnoth. It is on the Companies register there.
2. Not 50x. On one question Laya runs at about 33 ms on a T4 GPU against Jev at roughly 240 to 280 ms, so about 8x. The bigger multiples only appear when you batch questions or run it on an M3 Max, and Jev's numbers were never run head to head by the Laya team. They say so themselves.
3. About 1 GB is right, and yes it runs locally. That part is true and it is genuinely useful.
4. The catch nobody mentions: context. The English model takes 512 tokens and the others 1,024, with only about 320 to 770 of those for your data. A single CV is 1,500 to 3,000 tokens. Jev takes 32,000. Laya cannot read a CV, a contract or an email thread. It can only judge a short summary you have already extracted.
5. The accuracy win is narrower than the headline. It beats Jev on simple 4 or 6 way classification, but on 77 way choice it scores 0.43 against Jev's 0.87, and score questions land at 0.37. The typed-decisions result comes from a checkpoint fine-tuned on that test's own training set.
Laya is a good, free, open model that fits in a pipeline. It is not a drop in replacement for Jev, and it is not Chinese.
2026-09-21 17:40:35