Laya Review

Laya is a non-autoregressive "System 1" decision model published by Convai Innovations on Hugging Face and GitHub. You give it a state (a text, email, support ticket or JSON object) and a set of typed questions, and it returns typed answers with probabilities instead of generated text. Three question types exist: noul (yes/no), choice (pick one of your options) and score (an ordinal rating). Each option is scored at its own marker token and normalised across that question's options, so you can define new questions and labels without retraining, and the model cannot invent a category or return malformed output. Outputs include probabilities, prediction sets and abstain or review routing.

There are three checkpoints: an English model (ModernBERT-large backbone plus a decision head, 421M parameters, 512-token context), a multilingual model (mmBERT-base, 322M parameters, 1,024 tokens by default, up to 8,192 with accuracy varying past about 4,000), and a typed-decisions model fine-tuned on Convai's benchmark. A Router sends English text to the English checkpoint and everything else to the multilingual one. You install it with pip install laya (Python 3.10+, PyTorch on CUDA, Apple MPS, XPU or CPU), run laya-serve for an HTTP endpoint, or use the CLI; there are SDKs or clients for TypeScript, Go, Rust, Java, .NET and C++, integrations for LangChain, LlamaIndex and CrewAI, an MCP extra, a Hugging Face Space demo and a Colab notebook. An optional Vega engine adds image, audio and video input.

Laya is free and open source under the Apache 2.0 license; the model card lists its cost as "$0 self-hosted". Convai Innovations does not list a hosted, paid Laya API, so your only costs are your own hardware or cloud compute. For scale: the card's benchmarks ran on a single Tesla T4 GPU, CPU latency with preloading is 193 to 464 ms per call, and the downloads are about 808 MB (English) and 647 MB (multilingual). Fine-tuning recipes target Kaggle's free 2x T4 GPUs.

Prices checked on the vendor's model page on November 10, 2026. See the Laya model card.

What the Laya model card and repository list:

Laya suits developers who run high-volume classification, triage and routing (support tickets, intent, spam, moderation flags) and want fast, structured answers they can host themselves, including air-gapped, under a permissive license. It fits best when you can label a few hundred examples from your own domain: the model card says the base checkpoints are near chance on its typed-decisions benchmark zero-shot (0.362 accuracy) and reach 0.766 only after fine-tuning, that accuracy drops on questions with many options (0.425 on the 77-label Banking77 set), and that probabilities should be recalibrated on your own data.

If you want a hosted decision API that works without fine-tuning, compare TypeSafe's Jev, which the model card uses as its benchmark. For classic NLP pipelines with trainable text classifiers, see spaCy. If you need free-form reasoning or generated explanations, a general LLM API is the better tool.

User reviews are scarce; coverage so far comes from tech newsletters summarising the model card. i-SCOOP (September 29, 2026) credits Laya with fixed answer options that cannot be invented or malformed, Apache 2.0 weights that can run locally, and strong binary safety-style results, but calls it "a fast base to specialise" rather than a zero-shot engine, citing weak base-checkpoint accuracy, a big drop with many labels, fragile negation and over-confident probabilities. It suggests teams that need broad accuracy without fine-tuning may find a hosted option like Jev easier to start with. AI Weekly covered the release as a 421M ModernBERT decision model under Apache 2.0. All speed and accuracy figures above are Convai's own, measured on a T4; Jev's figures in the card are third-party published, not measured by Convai.