# Glimmer System One

- **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:** [Aryan Srivastava](https://cerebralvalley.ai/u/aryan-sentlyx)
- **GitHub:** https://github.com/aryan-sentlyx/glimmer-system-one
- **Demo video:** https://www.youtube.com/watch?v=xk8nNau3B6Y
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2690

Glimmer System One is an open decision model for the Muse family. It is Muse Glimmer 30B with the part that writes text removed. You give it a picture, a short state text and your questions. For each question it gives back one number: the probability that the answer is yes. One forward pass, 45 milliseconds per image, and the probability is calibrated.

The gap. Muse Spark is the reasoning model behind the API. Muse Glimmer is the open 30B multimodal model. What the family does not have is a System One model: a model that does not write an answer but decides, and returns a probability in one pass. TypeSafe Jev made that category popular in September. OpenAI followed with the Decisions API. Microsoft followed with Decision-1. Both are closed. The Muse family has none.

Why it matters. A production line moves one part per second. Muse Spark 1.3 writes a median of 1,144 reasoning tokens for one inspection image and takes 8.0 seconds. The part is long gone. Glimmer System One answers in 45 milliseconds. Because the weights are open, it can stop at layer 13 of 52 and answer in 21 milliseconds. It works without a picture too: it can stop an AI agent before a dangerous command, or route a support ticket and rate how urgent it is.

What you can ask. Three kinds of questions, in the request shape of the TypeSafe Jev Decisions API: is this true (a probability), which one of these (a probability per option), how much on a scale (the expected level and the distribution).

How it works. One batched forward pass over all the questions of a request. The zero-shot mode reads the logit of "yes" against "no" at the answer position and applies one fitted temperature. The head mode adds a 135 KB logistic head on the hidden state of layer 52, or layer 13 for the early exit, trained on half of the known products and scored on the other half.

Results on all 1,725 MVTec AD test images, 15 products, "does this part have a visible defect":
- Glimmer System One, head, held-out half: AUROC 0.962, accuracy 90.2%, calibration error 0.045, 45 ms.
- Early exit at layer 13: AUROC 0.954, calibration error 0.034, 21 ms.
- Liquid d1-ultra: 0.929, 0.147, 0.46 s. Closed.
- TypeSafe Jev 1.14: 0.917, 0.125, 0.31 s. Closed.
- Baseten Elephant 1: 0.902, 0.181, 0.43 s. Open.
- Muse Spark 1.3, JSON, default reasoning: 0.872, 0.101, 8.0 s. Closed.

The story is speed and calibration. When Glimmer System One says seventy percent, about seventy percent of those parts are defective. That is what makes a threshold mean something.

The release, all Apache 2.0: the code, the benchmark and the head training on GitHub; the heads, the temperature and the split on Hugging Face; a live demo page where you drop an image, type your questions as plain lines and press Run; a three minute video.

Code: https://github.com/aryan-sentlyx/glimmer-system-one
Heads: https://huggingface.co/aryan-sentlyx/glimmer-system-one-heads
Demo: https://aryan-59489--glimmer-demo.us-east.modal.run

## More from Meta Global AI Developer Hackathon

- [Spekios](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2687)
- [Antinoise](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2688)
- [Proof-Carrying-Code](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2689)
- [zaxyar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2691)
- [PodAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2692)
- [PRLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2693)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2690. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt