feat: add ondeviceml.space — browser-based on-device AI inference

[abhid1234]

What this adds

Registers ondeviceml.space as a local app for Gemma and Qwen2 transformer models.

ondeviceml.space is a client-side-only AI gallery: models run entirely in the browser via WebGPU / WASM (MediaPipe Tasks GenAI runtime). No server, no install, no account — the model weights download once to the browser cache and all inference stays on-device.

Deep-link behavior

When a user clicks "Open in ondeviceml.space" from a model page, the URL

https://ondeviceml.space/?hf_model=google/gemma-3n-E2B-it&task=image-text-to-text

lands on the Gallery page, which:

1. Matches the hf_model param against the local catalog
2. Shows a confirmation banner: [ Gemma 3n E2B → Chat ] [ Load & Open ] [ Dismiss ]
3. Downloads the model weights (first visit only; cached via browser storage after)
4. Navigates to the matched feature

The deep-link intake is live on production — you can test it now:

https://ondeviceml.space/?hf_model=google/gemma-3n-E2B-it&task=image-text-to-text
https://ondeviceml.space/?hf_model=Qwen/Qwen2.5-1.5B-Instruct&task=text-generation

displayOnModelPage logic

Shows only for transformers-library models with text-generation or image-text-to-text pipeline whose model_type is one of gemma3n, gemma4, gemma3, qwen2 — the architectures validated to run via MediaPipe in-browser. Happy to expand or tighten the list based on your feedback.

Catalog today

Model	HF pipeline_tag	Browser runtime
Qwen/Qwen2.5-1.5B-Instruct	text-generation	WebGPU/WASM
google/gemma-3n-E2B-it	image-text-to-text	WebGPU/WASM
google/gemma-4-E2B-it	image-text-to-text	WebGPU/WASM
google/gemma-3n-E4B-it	image-text-to-text	WebGPU/WASM

Checklist

• Deep-link is live and tested on production
• displayOnModelPage targets only supported model types
• No snippet needed — the deeplink opens a web URL directly
• No macOS-only flag — works on any browser with WebGPU support (Chrome 113+, Edge)

---

Note

Low Risk
Low risk: adds a new LOCAL_APPS entry and display gating logic plus an external deeplink URL; no existing execution paths or core inference logic are modified.

Overview
Adds ondeviceml.space to the LOCAL_APPS registry so supported Hugging Face model pages can show an “Open in ondeviceml.space” option.

The new entry is only shown for transformers models with text-generation or image-text-to-text pipelines whose config.model_type starts with one of gemma3n, gemma4, gemma3, or qwen2, and it deep-links to https://ondeviceml.space/?hf_model=...&task=....

^{Reviewed by Cursor Bugbot for commit fcba75d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit fcba75d. Configure here.}

[cursor]

Prefix allows unsupported architectures

Medium Severity

startsWith broadens the allow-list, so distinct architectures like qwen2_5_vl pass the qwen2 check. The button can appear for models outside the ondeviceml catalog or validated runtime, producing deeplinks the site cannot load.

 

^{Reviewed by Cursor Bugbot for commit fcba75d. Configure here.}