Layla

When building a software architecture with Layla, language detection must be completed before the forward pass and cannot rely on a fallback strategy of switching after observing low confidence post-inference.

Layla is not a single model but a model family comprising three checkpoints: English, multilingual, and typed decision.

Layla uses a bidirectional encoder architecture, replacing the traditional language head with a decision head composed of two Transformer layers, and a single forward pass outputs only structured decisions without generating text.

Layla's output primitives fall into three types: Choice (single-select, outputs the selected item and distribution probability), Score (rating question, gives the expected grade), and Bool (yes/no question, returns a single probability value).

The English version of Layla has 420 million parameters, the multilingual version has 320 million parameters and covers hundreds of languages, and there is an additional version fine-tuned for workflows.

The official claim is that multiple questions in a single call are completed within one forward pass, but the Anthros team and AXERA NPU deployment package tests found that each additional question triggers one additional forward pass.

The Layla authors never obtained API access to Jev and have never tested it themselves, and the claimed 7.8× speed advantage over Jev compares local single-GPU performance at maximum against cross-continental public network latency from someone calling the cloud API from France.

Running Layla on an ordinary i5 processor, a single question takes up to 416 milliseconds, and 50 concurrent questions freeze for 14.5 seconds, making the speed heavily dependent on local GPU compute.

After fine-tuning on the Typed Decisions dataset, Layla scores 0.766, but the untuned base model scores only around 0.35, which is lower than the baseline of 0.461 obtained by always guessing the majority option.

Layla's official claim of a calibration error of 0.081 outperforming Jev is the result after re-fitting the temperature; the native calibration error is actually worse than Jev's, and the multilingual version ships with an empty temperature table, causing probability collapse.

Layla is extremely sensitive to the number of options and their positions; on 77 classification tasks its accuracy drops to 0.425, and shuffling the option order causes 20% of answers to flip.

When the English version of Layla encounters unknown characters such as Khmer, its accuracy is 0, yet the average confidence is as high as over 95%, indicating that the model's confidence measure breaks down when encountering characters outside its script coverage.

The English version of Layla has a context length of 512 and the multilingual version 1024; inputs that exceed these limits are silently truncated without any error or warning.

Layla is an open-source model family released three days after Jev, using the Apache 2.0 license for its open-source weights.