Glimmer Scope

Glimmer Scope

Steer the Muse Glimmer 30B language model with features found inside its own layers. Every question returns two answers, one plain and one steered, so you see what each dial does.

Made by rob-x-ai on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Glimmer Scope lets you steer the Muse Glimmer 30B language model with features taken from inside its own layers. Every question you ask comes back twice. One answer is the model as it is, and the other is the model with a few features turned up.

Both replies are produced together under greedy decoding. That leaves the steering as the only thing that differs.

Who it suits: It suits researchers and students who study how language models work inside, and people who want to see feature steering at work on a 30B model.

Industry | Who uses it | How they use it | Example
AI research | Interpretability researchers | Compare a plain answer with a steered one | Seeing what a single feature from layer 24 does
AI safety | Safety researchers | Study how internal features change behavior | Checking how a character direction shifts a reply
Education | Students | Try a dial and read the two answers | Learning what steering a model means
AI development | Developers | Read app.py to see how the demo runs the model | Reusing the two-answer layout for another model

Skip it if you want an ordinary chat with a model, since each question returns a plain and a steered answer.

Where to use it: The app runs in these places.

• A web browser, on the Space's page at Hugging Face
• A local copy on Windows, macOS or Linux, running Python 3.12, if you have the hardware
• Hugging Face's zero-a10g hardware at its largest ZeroGPU size, where the online Space runs

Use case | Who | What they get back
Compare plain and steered answers | Researchers | Two answers to one question
Try a feature dial | Students | The effect of one internal feature
Run the interface without a model | Developers | The page alone, with no model loaded

How to set it up: Online you only need a browser. A local copy needs these.

• Python 3.12 on Windows, macOS or Linux
• The pip command
• A GPU with room for a 30B model at full precision, since the README sizes the online Space at 96 GB
• A Hugging Face token with read access, while the dataset it names is private

1. Install the huggingface_hub package.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Install gradio at version 6.31.0.
5. Install the packages listed in requirements.txt.
6. Start the app with python app.py.

pip install -U huggingface_hub
hf download rob-x-ai/glimmer-scope --repo-type space --local-dir glimmer-scope
cd glimmer-scope
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page, or run the command below in your copy and open the page the app serves. Then ask a question, pick the dials to turn up and run it.

python app.py

Four environment variables change how a copy runs. GLIMMER_DIR points the app at a different model folder, and GLIMMER_DATA does the same for the data folder. VIDEO_URL puts a demo video link on the page. Setting GLIMMER_MOCK to 1 runs the interface without a model.

Example: You ask one question and pick a feature dial. The app returns two answers, one from the model as it is and one from the model with that feature turned up.

For newcomers and experts: If you are new, open the Space's page and try one of the supplied examples. Reading the two answers side by side shows what a dial changes.

If you know the field, the dials are of two kinds. Some are single features taken from the autoencoders for layers 24 and 36. Others are character directions, found by setting in-character answers against plain ones. The README says each passed an automated steering test with its top grade, and that all concern harmless subjects. Greedy decoding in one batch keeps the comparison fair.

Ask a question and pick the dials
The model answers once as it is
The model answers again with features raised
You compare the two answers

Good to know: The online Space asks for the biggest ZeroGPU size, 96 GB, because the model is not shrunk by quantization. Two read-only mounts hold the model and the data, so nothing large downloads at startup. Without the data mount, the Space downloads its dials, examples and images instead.

• License: apache-2
• Cost to run: the large ZeroGPU size above. A token secret with read access is needed while the data repository is private
• Data it reaches: the Muse Glimmer 30B model, the Glimmer Scope autoencoders and a dataset of dials, examples and images
• Not for: casual chat or low-memory machines, since the model runs at full precision

On Hugging Face, 13 November 2026

• Likes: 2
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/rob-x-ai/glimmer-scope
• Files: https://huggingface.co/spaces/rob-x-ai/glimmer-scope/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.