Skip to content
Agentic Workflowdocs
v0.8.2Install free

use the app

Add a custom local model

Add an on-device model beyond the built-in list: an MLC-compiled WebLLM model, a TensorFlow.js model.json or a transformers.js model with an ONNX build.

On the Local AI page, click Add custom model. It takes two steps.

Step 1:The model

Pick the engine; the dialog asks for exactly what it needs:

  • WebLLM: an MLC-compiled model, with its weights URL and its compiled model library (.wasm). WebLLM models are always chat models, and WebLLM can’t load a raw .gguf: it needs the MLC build.
  • TensorFlow.js: the model URL of a graph or layers model (a model.json).
  • transformers.js: the Hugging Face model id (for example Xenova/vit-base-patch16-224). It needs an ONNX build (an onnx/ folder); repos with only GPTQ, GGUF or safetensors weights won’t run in the browser.
Step 1, close-up of the app: Engine, outlined.Step 1, close-up of the app: Engine, outlined.Full screen: the whole screenshot for step 1
Engine
Step 1 of 2 of Add a custom model: the Transformers.js engine is selected, the Hugging Face model id is Xenova/vit-base-patch16-224, and a green box shows what was detected from the model's config (task Image classification, input Image, 1000 labels, ONNX · about 84 MB), above Use it as, Display name and Continue.Step 1 of 2 of Add a custom model: the Transformers.js engine is selected, the Hugging Face model id is Xenova/vit-base-patch16-224, and a green box shows what was detected from the model's config (task Image classification, input Image, 1000 labels, ONNX · about 84 MB), above Use it as, Display name and Continue.

Check: the fields for that engine are filled in.

Step 2:Check the task, then Continue

As you type, the dialog detects the model’s task and shows what it found: task, expected input and output, file size where known, and where the guess came from. For transformers.js it reads the model’s hub metadata; for a TensorFlow.js model.json it reads the input and output shapes (for example [1,224,224,3] → [1,1001] means image classification). If the guess is wrong, or nothing was detected (for example offline), choose the task under Use it as. The task decides which nodes can pick the model. Then click Continue.

Step 2, close-up of the app: Use it as, outlined.Step 2, close-up of the app: Use it as, outlined.Full screen: the whole screenshot for step 2
Use it as
Step 1 of 2 of Add a custom model: the Transformers.js engine is selected, the Hugging Face model id is Xenova/vit-base-patch16-224, and a green box shows what was detected from the model's config (task Image classification, input Image, 1000 labels, ONNX · about 84 MB), above Use it as, Display name and Continue.Step 1 of 2 of Add a custom model: the Transformers.js engine is selected, the Hugging Face model id is Xenova/vit-base-patch16-224, and a green box shows what was detected from the model's config (task Image classification, input Image, 1000 labels, ONNX · about 84 MB), above Use it as, Display name and Continue.

Check: Use it as shows the task you want.

Step 3:Confirm you trust the source

Tick I trust this source and want to add the model. Then click Add & download.

Step 3, close-up of the app: I trust this source, outlined.Step 3, close-up of the app: I trust this source, outlined.Full screen: the whole screenshot for step 3
I trust this source
Step 2 of 2 of Add a custom model, Only add models you trust: Xenova/vit-base-patch16-224 (Transformers.js), a warning that the model isn't reviewed by awflow and is downloaded from huggingface.co, the I trust this source and want to add the model checkbox, and Back and Add & download (84 MB) buttons.Step 2 of 2 of Add a custom model, Only add models you trust: Xenova/vit-base-patch16-224 (Transformers.js), a warning that the model isn't reviewed by awflow and is downloaded from huggingface.co, the I trust this source and want to add the model checkbox, and Back and Add & download (84 MB) buttons.

Check: the model is listed with a Custom badge and starts downloading.

Once added, a custom model gets a Custom badge and behaves like a built-in one: install, test, use it in a workflow (including the Run Local Model node, whose Custom model (raw output) task runs a model of any other kind), and remove it, which deletes both its weights and the entry.

Ask Aria