What is Laya? The open-source Jev alternative, and when to train your own model

Laya is a self-hosted alternative to a hosted decision API. For the hosted approach, read Jev vs a trained classifier.

From its model card and README:

• Three checkpoints: English and typed-decisions checkpoints use ModernBERT-large (421M parameters); the multilingual checkpoint uses mmBERT-base (322M). The router chooses a checkpoint for the input, or you can select one explicitly.
• Three question types, the same as Jev's: a choice from your options, an ordinal score, and a yes/no probability.
• It never writes text. It returns structured answers, but valid structure does not guarantee the category is correct.
• Fast and small. The project reports 32.8 ms for one multilingual question and 39.5 ms for English on a Tesla T4 GPU, and a few hundred milliseconds on a CPU. The downloads are under 1 GB.
• Ways to run it: a Python package (pip install laya), a self-hosted HTTP server that matches Jev's API shape, and add-ons for MCP, LangChain and ONNX.
• Licence: Apache 2.0, so you can run it commercially, on your own hardware.

In short: Jev is TypeSafe's hosted API, and doesn't train on your data. Laya is open source, runs on your own hardware, accepts Jev's request format, and can be fine-tuned. Both return a choice, a score or a yes/no with a confidence. Which is more accurate depends on the task, and they haven't been measured in the same test.

The full side-by-side, including the published Banking77 figures and where they come from, is on Jev vs Laya vs a trained classifier.

Laya is open source; the infrastructure is yours to operate. You can use its Python package or host its Jev-compatible HTTP server on your own hardware.

Plan for compute capacity, model loading, monitoring and server security. The server has no authentication unless you configure an API key. Keeping inference on your infrastructure gives you control over deployment, but your team still handles access, updates and data protection.

The model is free under Apache 2.0. Hardware and the time to maintain it are part of the cost. See the project’s self-hosting instructions.

The model card is frank, which is useful:

• Zero-shot, the base checkpoints are close to chance on its typed-decision test. The project reports 0.362 for its base English checkpoint and 0.766 for a checkpoint fine-tuned on that benchmark’s training split.
• Many options are hard. The 77-label Banking77 scores 0.425.
• It ships over-confident. The card recommends refitting its confidence on your own data: the calibration error drops from 0.466 to 0.081 after that.
• The English checkpoint confidently fails on non-Latin scripts. The router is there to prevent that.
• The self-hosted server has no authentication unless you set an API key.

These are project-reported results on specific tests. They show the importance of training it on your own labeled data, and checking it on your data.

Yes. The project has a fine-tuning notebook that runs on two T4 GPUs. That's the same idea as a trained classifier: learn from examples your team has already labeled.

What you take on when you do it yourself:

• preparing the data, cleaning duplicates and thin categories, and removing personal information;
• running the training;
• holding back test examples, and measuring accuracy per category;
• calibrating the confidence;
• hosting the model and securing its server;
• building a way for people to review uncertain answers, and to feed corrections back.

StayCharted manages training, evaluation and review in a browser. Your team still prepares representative data, approves labels and decides whether the result is ready to use:

• Train on your examples: upload what your team already decided, or, on Essentials and Business, start from unlabeled rows with suggested labels you confirm.
• Check your data first: duplicates, categories that mean the same thing, thin categories, and personal information, with tools to review the findings; detection is not exhaustive.
• See how accurate it is: every version is tested on examples it never saw, overall and per category. On Banking77, StayCharted's trained models answered 92.6% (AI Classifier Model) and 93.7% (Dedicated AI Model) correctly on held-out rows (the banking benchmark). That's a different setup from Laya's zero-shot figure, so it isn't a head-to-head test.
• Review (Essentials and Business): answers below your cutoff wait for a person, and corrections train the next version.
• Use it anywhere: files, Google Sheets, the API, Zapier, Make, Power Automate, Claude and ChatGPT. Pictures too, with Vision on Essentials and Business.

Choose Laya when:

• you have developers, and a GPU or capable CPUs;
• you want open source and full control;
• the decisions are generic, or you're prepared to fine-tune and calibrate it yourself.

Choose a managed trained model when:

• business teams must build and improve it;
• you need measured, per-category accuracy and a review loop;
• you'd rather not host anything.

Checked October 20, 2026 against the project’s model card and GitHub README, plus TypeSafe’s model documentation. Vendor results are not a shared head-to-head test.

• Jev vs Laya vs a trained classifier
• Training best practices
• StayCharted plans