最新模型
Laya releases non-autoregressive decision models, with multilingual version taking about 33 ms per question on a T4
Laya uses a bidirectional encoder to produce option probabilities directly and has released weights, an SDK, and evaluation materials. The author's tests show potential for low latency, but Traditional Chinese accuracy, probability calibration, and model-switching costs still require separate validation.
ConvAI Innovations published a technical overview of its Laya decision models on September 18, with releases of the Python package beginning the same day. The project provides open weights for producing numerical results directly for classification, ordinal scoring, and true/false judgments, supporting uses such as customer service triage, document processing, and agent routing. [Author's overview](https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me), [Package history](https://pypi.org/project/laya/)
The core architecture combines a bidirectional encoder with a decision head. Each candidate option is assigned a mask token; after scoring, softmax is applied across the options for the same question. The English model has approximately 421 million parameters, while the multilingual model has approximately 322 million. Callers can define options at request time without requiring the model to generate an answer token by token. This can reduce text parsing and formatting errors, although incorrect classifications and overconfidence remain possible. [Model card](https://huggingface.co/convaiinnovations/laya)
Training uses RLCD, rewarding probability distributions through a strictly proper scoring rule and updating the model through policy gradients. In theory, this objective encourages truthful probabilities, but it does not guarantee good calibration in a new domain. The model card explicitly calls for recalibrating the temperature parameter using your own data. It also acknowledges that the general-purpose weights perform worse than the majority-class baseline in some workflows; the higher scores come from specialized fine-tuned versions. [Architecture and limitations](https://huggingface.co/convaiinnovations/laya)
In the author's measurements, the multilingual model took about 32.8 milliseconds to process a single question on a T4, and 72.3 milliseconds for a batch of ten questions. Deployment also involves a cold-loading cost: the router keeps only one model loaded by default, so alternating between languages may trigger reloading. The documentation reports a median reload time of 10.3 seconds on a T4. Preloading multiple sets of weights can reduce switching delays, but teams must assess the resident memory requirements themselves. [Deployment notes](https://pypi.org/project/laya/)
Results for Traditional Chinese are already available for inspection. In a MASSIVE intent classification test with 20 options, the author reports 54% accuracy and an expected calibration error of 0.327 for the multilingual model on `zh-TW`. This demonstrates measurable Chinese-language capability while also revealing discrepancies in its probability estimates. Large numbers of options also divide the decision head's token budget, and options may become difficult to distinguish when their text is truncated. [Public benchmarks](https://github.com/NandhaKishorM/laya/blob/main/BENCHMARKS.md)
For engineering teams, the open weights and evaluation code provide a starting point for building a fast local classifier. The next step is to use Traditional Chinese business data to check accuracy, abstention thresholds, and sensitivity to option order, while including cold loading in latency tests. Traffic spanning multiple languages also needs separate load testing so that averages do not conceal switching pauses. The project's comparisons with Jev use different sources, prompts, and samples, so they cannot yet be treated as a fair head-to-head determination of a winner. [Comparison conditions](https://github.com/NandhaKishorM/laya/blob/main/BENCHMARKS.md)