Laya
An open-source non-autoregressive decision model based on the ModernBERT-large encoder with 421 million parameters. Trained using the RLCD reinforcement learning method, it is designed for decision-oriented tasks such as intent routing, content moderation, and fact-checking.
Core Facts
Timeline (last 90 days)
DeciFlow uses containerized deployment, the repository contains Docker instructions for the DeciFlow UI and the Laya inference server, and the browser interface runs on port 4000 by default
DeciFlow nodes can monitor logs locally while sending inference requests to a remote shared Laya server, and the log source machine does not need a GPU or a PyTorch installation
DeciFlow supports two backends: Laya is the default option, supports self-hosting, and can run on CPU or CUDA; Jev is an alternative hosted decision API
In the experiment of swapping option names and definitions, when Laya used yes/no naming, 76.9% of decisions flipped, and the original paired AUC of 93.8% dropped to 23.2% for the swapped pairing.
After replacing option names with random five-character strings, all three models returned to the flip rate of neutral names, indicating that meaning is carried by the definitions rather than the names.
Polarity matters more than familiarity: even pairings that Laya never output during training (such as absent/present) still raised the flip rate by 41 percentage points compared with neutral names.
Sun, Xu, Shi, and Yang of Fudan University and the University of Science and Technology of China tested three models, Jev, Laya, and Open-Jev, in the paper "Type-Safe Is Not Error-Free" (September 2026).
Laya is an open-weight model with the same form as Jev that can be self-deployed and run, which to some extent limits vendor lock-in risk.
The Laya and Clef decision models can now be trained on Apple Silicon, and Clef also supports image training.
The same Laya weights scored 55 on llama.cpp but only 41 on PyTorch and MLX; changing the JSON state to labeled text raised the Laya judge group score from 1/10 to 4/10.
40 more timeline events
All Facts (20)
On a T4 GPU, Laya takes 33 milliseconds per single question, and batch processing can compress that to 7 milliseconds
65%UnverifiedLaya is an open-source, open-weight classification decision model that can be downloaded and run locally.
60%UnverifiedSmall decision models can serve as front-end schedulers or decision nodes in large model systems, playing a key role in Agent workflows and multi-model orchestration architectures.
60%UnverifiedLaya pairs a bidirectional ModernBERT-large encoder with a Transformer head trained from scratch that scores [MASK] option tokens.
60%UnverifiedThe entire Laya model was trained on a single RTX 6000 Pro (96 GB VRAM).
60%UnverifiedRLCD is a policy-gradient-based reinforcement learning method (an unofficial name coined by the author) that optimizes decision models for strictly proper scoring rules, granting maximum reward only when the model outputs mathematically calibrated probabilities.
60%UnverifiedDeciFlow uses containerized deployment, the repository contains Docker instructions for the DeciFlow UI and the Laya inference server, and the browser interface runs on port 4000 by default
50%UnverifiedDeciFlow nodes can monitor logs locally while sending inference requests to a remote shared Laya server, and the log source machine does not need a GPU or a PyTorch installation
50%UnverifiedDeciFlow supports two backends: Laya is the default option, supports self-hosting, and can run on CPU or CUDA; Jev is an alternative hosted decision API
50%UnverifiedPolarity matters more than familiarity: even pairings that Laya never output during training (such as absent/present) still raised the flip rate by 41 percentage points compared with neutral names.
50%UnverifiedAfter replacing option names with random five-character strings, all three models returned to the flip rate of neutral names, indicating that meaning is carried by the definitions rather than the names.
50%UnverifiedIn the experiment of swapping option names and definitions, when Laya used yes/no naming, 76.9% of decisions flipped, and the original paired AUC of 93.8% dropped to 23.2% for the swapped pairing.
50%UnverifiedLaya is an open-weight model with the same form as Jev that can be self-deployed and run, which to some extent limits vendor lock-in risk.
50%UnverifiedThe same Laya weights scored 55 on llama.cpp but only 41 on PyTorch and MLX; changing the JSON state to labeled text raised the Laya judge group score from 1/10 to 4/10.
50%UnverifiedBoth the Clef model and the Laya model with a Qwen3.5 backbone can be exported to GGUF format and served via llama.cpp, including multimodal decision models capable of understanding images.
50%UnverifiedThis version supports exporting Clef models based on the Qwen3.5 backbone, as well as Laya models, to GGUF format.
50%UnverifiedWhen faced with 50 nearest-neighbor tools with a single correct tool present, Laya's accuracy drops sharply to 31%, while Jev maintains 98%.
50%UnverifiedOn the RAG relevance gating task, Laya and Jev tied, with neither demonstrating a decisive advantage.
50%UnverifiedWhen option order is reversed, the open-weight model Laya changes 30% of its answers.
50%UnverifiedThe hosted model Jev is significantly more accurate on 9 out of 11 decision points, with margins ranging from +10.8 to +46.0 percentage points.
50%