Open-Source 421M Discriminative Model Laya Goes Viral: 7x Faster but Falters on Complex Tasks

An independent developer released the open-source model Laya (421M parameters, Apache 2.0) on September 18, scoring 1200+ points and roughly 300,000 views on Hacker News within a day. The model returns to the discriminative paradigm: given input text and a typed question, it directly outputs a probability-scored answer in a single forward pass, generating no text at all, with latency of about 40 milliseconds on a single T4. The author also published a blog post, "Laya, Jev and the Return of the Discriminative Model," exploring the comeback of this small discriminative approach and comparing it with the closed-source product Jev (from TypeSafe AI).

Confirmed

Laya is a 421M-parameter open-source model under Apache 2.0 that returns a probability-scored answer in a single forward pass, runs at about 40ms on a T4, and generates no text.
The author's benchmarks: 0.950 accuracy on AG News with roughly 7x faster speed; on 77-class Banking77, only 0.425 versus a baseline of 0.870.
The model card acknowledges that its base checkpoint underperforms most baselines in zero-shot settings.
The author's companion blog post argues that the discriminative small-model approach is making a comeback, and compares it with the closed-source Jev.

Why it matters

With generative LLMs dominating today, Laya shows the latency and cost potential of small models that skip text generation and directly output probabilistic answers.
However, its large lag on multi-label/multi-class tasks reveals clear limits to this approach, with the base checkpoint's zero-shot weakness being the main shortcoming—something the author openly admits in the model card.

2026-09-28 ~ 2026-09-28 · 5 related posts

Primary sources

421M-parameter open-source Laya returns answer probabilities in one pass at 40ms on a T4 — maier_ak · 2026-09-28 12:30 (UTC+08:00)
Laya hits 0.950 on AG News and is 7x faster, but flops on Banking77 — maier_ak · 2026-09-28 12:30 (UTC+08:00)
421M-parameter discriminative model Laya is 7x faster but fails badly on 77-label tasks — maier_ak · 2026-09-28 12:30 (UTC+08:00)

Laya and Jev: the return of the discriminative model, fast but narrow — maier_ak · 2026-09-28
[source] 421M-parameter open-source Laya returns answer probabilities in one pass at 40ms on a T4 — maier_ak · 2026-09-28
Laya: a 421M open-source model answers with probabilities in one 40ms forward pass — maier_ak · 2026-09-28
[source] 421M-parameter discriminative model Laya is 7x faster but fails badly on 77-label tasks — maier_ak · 2026-09-28
[source] Laya hits 0.950 on AG News and is 7x faster, but flops on Banking77 — maier_ak · 2026-09-28