Deadbug Local Cloud β the IMP-VISION host rig
Deadbug Local Cloud is the self-hosted half of IMP-VISION: the glasses' phone app and its LLM Brain, running on a box we control instead of inside a phone session. The phone becomes a thin client β browser, camera, mic β and the models live server-side, next to the server that runs them.
Why it exists
On-phone inference works until it doesn't: mobile operating systems kill long downloads and background model servers without asking. Local Cloud removes that whole failure class:
- Models download once, server-side, at datacenter speed.
- llama-server runs as a service β it is never swiped away.
- The phone needs no local models at all: open the rig's address, Ask, use the Senses feed from the phone's own camera.
The stack
| Layer | Choice |
|---|---|
| App | IMP-VISION app (Flask, phone-first UI) |
| Brain host | llama.cpp llama-server, GGUF Q4_K_M |
| Default Brain | Qwen/Qwen2.5-1.5B-Instruct-GGUF (Apache-2.0) |
| Model sets | Devilishcode/imp-vision-model-sets |
| House Brain (in training) | Devilishcode/deadbug-brain-qwen2.5-1.5b |
| Access | HTTPS tunnel, gated β the rig is private by default |
The rig is hardware-agnostic: this stack runs on a cloud VM today and moves to any Linux box (a salvaged mini-PC is the house style) by copying one folder and running one start script.
Links
- Code + setup: github.com/deadbuglabs/local-cloud
- The glasses: github.com/deadbuglabs Β· Deadbug Labs on Facebook
- Live demo of the picker concept: IMP-VISION Space
Built by Deadbug Labs β boards and the tools that run on them, designed in the open.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support