# Ultimamind

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aliaksandr Kamovich](https://cerebralvalley.ai/u/abrasex)
- **GitHub:** https://github.com/aim-do/glimmer-decide
- **Demo video:** https://youtu.be/DawY6e_nZxo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3407

Glimmer Decide is a camera test bench for open-weight decision models, and a way to train your own. You point a phone or webcam at a desk and say what you want to do (“cut the paper along a straight line”). The model picks one object from a closed set (scissors, stapler, tape, marker and so on) or refuses with “nothing here fits” or “show me another angle”, with a probability for every option. SAM 3.1 then highlights the chosen object on the frame.

It solves two problems. The first is speed. A generative model reasons in text before it answers: Muse Glimmer with reasoning took 3–9 seconds per question in our tests, too slow for glasses or a live camera assistant. Decision models answer in a single forward pass with no generated text: JevAny answered in about 0.7 s end to end, MetaEncoder in about 1.7 s. But open-weight decision models are new, were trained mostly on text, and there was no easy way to test them on real camera frames. Glimmer Decide runs them side by side on the same frame with their timings. That lets us try features quickly and keep what works: refusing, asking for a better view, a camera-off ablation, highlighting, phone capture.

The second problem is that a released model does not know your scene. On an empty grey frame, JevAny still chose scissors with p=0.88 from the goal text alone. So the same page collects training data: every saved frame with its right answer is a training example. One command then fine-tunes JevAny on your frames, or trains a decision head on Glimmer from scratch, on your own GPU.

## More from Meta Global AI Developer Hackathon

- [mountain-dew-esports](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3404)
- [V.E.R.S.O.](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3405)
- [RealityForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3406)
- [Civica](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3408)
- [FarmOps AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3409)
- [Evidient](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3410)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3407. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt