glance-speedlab
by yoheinakajima
Glance Speedlab is a browser-based test bench for measuring how quickly a local vision-language model can answer questions about a live camera feed. It records timing for each stage and documents which speed-up ideas worked and which did not.
- Stars
- 53 +2 a day, est.
- Licence
- MIT
- Runs on
- Your computer
- Setup
- Technical
- Language
- Python
- Age
- 34 days
What you can do with it
- Run a live camera feed through a local vision model and see capture, transport and total latency separately
- Compare the PyTorch reference backend with the faster 8-bit MLX scorer side by side
- Read the published findings to decide which speed-up techniques are worth trying in your own vision project
Make money with it
A freelancer could offer paid prototyping or benchmarking of on-device camera AI for clients who want to keep video data local.
An idea, not financial advice. Check the licence before charging for anything built on it.
Ways a business could use it
- Test whether a local camera-monitoring idea, such as checking a shop counter or workspace, is fast enough before building it, without paying for cloud vision APIs
- Use the benchmark results to choose a model size and settings for a privacy-friendly on-device vision prototype for a client
Can you use it commercially?
Free to use, change and sell, including in paid products. Keep the copyright notice.
Not legal advice. Read the project's licence file before building a product on it.
Getting it running
You clone the Glance repository, sync it with uv and cache the model, then install this repository with pnpm and start it with "pnpm launch" (or "pnpm launch:mlx" on Apple Silicon).
Follow the official instructions in the project README.
Things to check
- Research prototype, not a production monitoring system
- Measurements cover only one Apple M5 laptop and a small test suite
- Needs a source checkout of Glance and several gigabytes of model downloads
- MLX path works only on Apple Silicon and is experimental
Who's talking about it
Checked 28 Oct 2026. Summary drafted by AI from the project's README.
