v1.7.0 Release Notes
We’ve released our own AI model, Glimpse-v1! While general multi-modal LLMs can describe images well, they often miss the contextual intent behind a scene. Sending every motion event to a cloud LLM is slow, costly, and exposes private footage to third parties.
We solved this by post-training Gemma 3 with reinforcement learning on thousands of labeled samples. Through this process, our model has learned that, for example “a person wearing a blue uniform carrying a package, with a gray van parked near the driveway”, is actually an “Amazon delivery”. Glimpse also understands the difference between a false positive (triggered by motion), and an actual event.
Additionally, the model was trained to generate event title and description in a single call. This significantly reduces inference time, so notifications can be sent more quickly.
We believe privacy is fundamental to smart homes. As such, we encourage everyone to run the AI models needed for LLM Vision on a local machine. By building specialized, compact models that can run locally on hardware with limited memory and compute resources, we aim to make local AI accessible to everyone.
While we’re excited what’s possible in the future, this is the first release of our model. As such, there will be limitations. For example, the model can currently only generate English responses.
We need your help! Models like these need a lot of training data to understand different situations well. You can help make Glimpse better for everyone by submitting feedback through the LLM Vision Card.
Check out Glimpse-v1 on our website: https://llmvision.org/glimpse/
Feedback is entirely optional, and never happens without your explicit permission. We do not collect any personal data. See our privacy policy.
Contributors
A huge thank you to our contributors @teofanis, @Dinnsen, and @radzio!
✨ Features
- You can now provide feedback for events to help us improve Glimpse.
🌐 Languages
🔧 Improvements & Fixes
- Set min height: The card now has a fixed
min-heightwhen used on a dashboard without layout. (by @teofanis) - Fixed accent characters: Translations for Polish, Italian and French used incorrect characters. Translations now use the proper accent characters for these languages. (by @radzio)
- Preview card icon: the preview card now renders the correct icon based on the category and label of the event.
- Offline support: All cards now come with bundled dependencies, so no internet connection is needed anymore. (by @valentinfrlch) (#105)
- Improved default colors: New default colors to identify events more intuitively based on category.
- Filters: Filters have been moved to the backend and are now directly integrated into the API. (#30, #96)
