Run ONNX language models in your browser

Use a supported ONNX language model for chat directly in your browser. Naidan’s Transformers.js manager downloads or imports its files and loads them from browser storage. This option is experimental and available in the Hosted edition; Standalone shows it but disables model actions.

First, open /#/settings/transformers-js to reach the Transformers.js model manager.

If first-run setup appears, close it with the × at the top right. The linked Transformers.js settings tab stays open for the steps below.

A browser that does not support or allow OPFS also disables model actions. Model support and performance depend on the model, browser and device.

1. Open the Transformers.js manager

In the Hosted edition, check that the manager offers Add New Models. If OPFS is unavailable, the manager explains the restriction and disables model actions.

The Hosted manager opens with model actions available when OPFS works.

2. Download a model and let it load

Choose one of the preset model paths or enter a Hugging Face model ID, then select Download Model. Local-only cannot download from Hugging Face; instead, use Select Model Folder to import an ONNX model folder obtained separately. Naidan explicitly downloads and saves the required files when the download is allowed. When that succeeds, the same action continues into a separate load from the local cache. The Engine Control panel shows when the model is ready. A listed model is only an example; choose one your setup supports.

Download Model completes file preparation, then starts a local cache-only load.

3. Reuse a downloaded model later

Under Downloaded Models, select Load for a complete model to load its local files again. An Incomplete entry offers Resume, which explicitly continues its download and then loads it. You can also import a folder of ONNX model files, after which it appears in the local list. Loading a complete model does not repair missing files by contacting Hugging Face.

The list distinguishes a complete model’s Load from an incomplete model’s Resume.

4. Save it as the chat default

Return to Connection in Settings. Set API Provider to Transformers.js, choose the downloaded model in Default Model, and select Save Changes. This saves the chat default; the manager’s Ready status by itself does not change the saved endpoint or model. This browser provider has no Endpoint URL field.

The saved Connection choice supplies the model to chats that inherit the global default.

When loading fails

Use the model manager to inspect the downloaded entry and any error. Load reads local model files; a missing required file is a load error. Resume is the explicit download path for an incomplete entry. The OPFS availability message can also explain why model actions are disabled. In Standalone, local-file restrictions affect the workers and WebAssembly required by this option.