Local AI · models that run in your browser

Private by physics, not by policy.

Models run inside your browser, on your hardware — there is nowhere for your text to go. No key, no bill.

With on-device models, 0 bytes leave this device

Add to Chrome — free What can it do?

The engines

Three runtimes in the tab. Two more nearby.

AWFlow picks the engine for each model. You never configure a server — the browser is the server.

01
WebLLM
Language models · WebGPU
Chat and text-generation models like Llama, Qwen and Phi, running on your graphics card through WebGPU.
02
Transformers.js
ONNX · vision, text, audio
The widest range of tasks: classification, captions, object detection, summaries, translation, embeddings and Whisper speech-to-text.
03
TensorFlow.js
Vision and embeddings
Vision and embedding models like MobileNet and COCO-SSD, on WebGPU with a WebGL or WASM fallback.

Chrome built-in AI
Gemini Nano, shipped with Chrome. Nothing to download in AWFlow. Not available in Firefox.
Ollama
Already running models on your computer? AWFlow finds them and lists them in the pickers.

What you can do on-device

Read, listen, see — without the cloud.

WebLLM
Chat
Ask, draft and rewrite with a private model.
- Llama
- Qwen
- Phi
Transformers.js
Summarise
Condense long text on your device.
- DistilBART
Transformers.js
Classify
Sort text into your own labels.
Transformers.js · TF.js
Embeddings
Power on-device knowledge search.
Transformers.js
Image captions
Describe what a picture shows.
- ViT-GPT2
Transformers.js · TF.js
Object detection
Find and box things in images.
- COCO-SSD
Transformers.js
Speech-to-text
Transcribe audio with Whisper, offline.
- Whisper

Model names are examples; availability depends on your browser and device.

Device fit

Honest about your hardware.

Before you download anything, the model manager shows how well each model fits your device and tells you plainly.

WebGPU check
Language models need WebGPU. Where your browser or graphics card can't offer it, the model says Not supported here.
Size before download
Every variant shows its estimated size up front, so nothing surprises your disk or your data plan.
Exceeds memory
If a model is likely too big for your memory, it says so. You can still install it, with eyes open.

Small chat model

Fits comfortably

Medium chat model

Tight fit

Large chat model

Exceeds memory

Illustrative — real sizes and verdicts appear in the model manager.

The model manager

Install, compare, remove in a click.

See every model, grouped by task. Compare up to three side by side, measured on your own device, and free up space when you're done.

The Local AI model manager, grouped by task

When you want more

Bring your own key. Pay the provider, not us.

For bigger jobs, connect your own provider with your own API key, or a model server on your computer. Your browser calls it directly; AWFlow never relays the request.