Open Source AI

Open-Source 421M 'Laya' Decision Model Surpasses Jev Benchmarks Using RLCD

Developer NandhaKishor has released 'Laya', a 421M-parameter non-autoregressive open-source decision model trained on a single RTX 6000 Pro. Pairing a bidirectional ModernBERT-large encoder with a scratch Transformer head, Laya scores mask option markers in a ~35 ms forward pass and surpasses all Jev benchmarks.

The AI Multiverse Editorial Desk•Friday, September 18, 2026•2 min read

Reported viaReddit LocalLLaMA (Model Leaks & Local AI)

Reader:

Executive Brief & 3-Point Takeaways

Laya is a 421M-parameter non-autoregressive decision model built on a bidirectional ModernBERT-large encoder.

Trained exclusively on 25,000+ human-annotated examples covering intent routing, moderation, and fact-checking.

Achieves rapid ~35 ms inference times and runs efficiently on low-end consumer hardware.

Impact:

Laya demonstrates that highly efficient, non-autoregressive decision models can achieve top-tier performance on complex routing, moderation, and scoring tasks while running easily on low-end hardware.

Fact-checked against primary sourcesCategory: Open Source AI

Open-Source 421M 'Laya' Decision Model Surpasses Jev Benchmarks Using RLCD

Illustration: AI-generated

Key Findings & Takeaways

Laya is a 421M-parameter non-autoregressive decision model built on a bidirectional ModernBERT-large encoder.
Trained exclusively on 25,000+ human-annotated examples covering intent routing, moderation, and fact-checking.
Achieves rapid ~35 ms inference times and runs efficiently on low-end consumer hardware.

Why It Matters

Key Details

Following community support for the Jev architecture, developer NandhaKishor M has released Laya, an improved open-source horizontal decision model trained on a 100% human-annotated corpus of over 25,000 real-world examples. The dataset spans intent routing, fact-checking, moderation consensus, prompt guardrails, rubric scoring, and multi-turn conversation trajectories without synthetic data shortcuts.

Technical Architecture

Laya features a 421M-parameter non-autoregressive architecture that pairs a bidirectional ModernBERT-large encoder with a scratch Transformer head. It evaluates and scores `[MASK]` option markers to resolve typed schemas in a single ~35 ms forward pass. The training utilized a policy-gradient reinforcement learning approach (RLCD) optimized against strictly proper scoring rules to ensure mathematically calibrated output probabilities.

What to Expect

Because of its compact size, Laya can run smoothly on consumer hardware and low-end PCs. Developers can test the live model via the Hugging Face Space (`convaiinnovations/laya-demo`), inspect the source code on GitHub, or download the model weights from the Hugging Face repository.