See inside
your model.
Run a language model on your device. Inspect its predictions and test how changes to its internal activity affect its replies.
Open source. Inference and saved work stay on your device. What leaves your device
For researchers, students, educators, and anyone curious about how language models work.
We’d love for you to feel at
Test what shapes a reply.
Ask a question, then try it with different response controls. Compare the predictions and replies in your browser, with no notebook or server to set up.
Read every token
See which words the model could have chosen at each step, with each token's probe and feature readings. The R and S lenses show what every layer is about to say and what it has read.
Add an influence
An influence nudges the model's internal activity toward a concept, word, or feature while it writes, or pushes a single neuron or silences a neuron or head. Strength sets how hard it pushes. The model's weights never change.
Compare what changes
Branch from any token in the Loom, Drowse's tree of continuations, and compare the new reply with the original. Your work autosaves in this browser.
Loading the recorded example…
Choose a model.
Use a chat model for conversation or a base model for text completion. Drowse checks what your device can run when you open the app.
Chat models
Follow instructions.
Explore a conversation.
Base models
Continue raw text.
Study next-token predictions.
Gemma 3 270M runs at full precision, so its download is larger than Gemma 3 1B's.
What you need.
A browser with WebGPU
Current Chrome, Edge, Brave, or Arc, or Safari 26 or later. Firefox support is a preview.
Storage
0.3 to 2.6 GB per model, plus optional insight packs. Your browser keeps them, so each model downloads once.
Memory
About 4 GB for the smallest chat model, up to 16 GB or more for the 4B models.
Phones
Phones run compact models as a preview. Use a computer for the larger models.
Words you'll see.
- Influence
- A change Drowse adds to the model's internal activity while it writes, pulling replies toward or away from a concept. Also called steering.
- Strength (α)
- How hard an influence pushes. 0 leaves the model unsteered; negative values push the other way.
- Word insights (R lens)
- A lens that shows which words each layer is leaning toward saying, before the model makes its final prediction.
- S lens
- A lens that shows which words each layer looks like it has read. It is the faithful read early in the model, the R lens late.
- Neurons and heads
- The model's own units. On Gemma 3 and Qwen3 models you can see which ones pushed a token, read what one writes, and push a neuron or silence either.
- Feature insights (SAE)
- A sparse autoencoder splits the model's activity into features: learned patterns that often match a recognizable idea.
- Loom
- A tree of continuations. Branch from any token, try another path, and compare the results.
- Base model
- A model trained only to continue text, without chat tuning. It predicts what comes next instead of following requests.
Common questions.
Does anything leave my device?
Models and optional packs download from Hugging Face. Some feature descriptions are fetched from Neuronpedia when you look them up; that request names the feature, not your text. Your prompts, chats, and saved work stay in this browser. The contact form sends only what you enter in it. Read the privacy details.
Why is the download large?
A model is mostly its weights: hundreds of millions to billions of numbers. Each model is 0.3 to 2.6 GB, plus any insight packs you choose. Your browser keeps the files, so you only download a model once.
Will it work on my phone?
Phones with WebGPU can run compact models as a preview, within the memory the browser allows. For the larger models and insight packs, use a computer.
Is steering the same as fine-tuning?
No. Fine-tuning trains the model further and changes its weights. Steering adds an influence to the model's activity while it writes. Set the strength to 0 and you get the unsteered model back.
What's a base model?
A model trained to continue text, before any chat or instruction tuning. It doesn't answer requests; it predicts what comes next. That makes it useful for studying raw next-token predictions.
Can I use it in a class?
Yes. Drowse is free and open source, and it needs only a browser. Each student downloads models to their own device, so check the storage and memory needs above first. Each model's license still applies.
How do I cite it?
Cite the Drowse version and the exact model checkpoint you used, alongside the methods Drowse builds on. The Credits page lists them.
How is this different from TransformerLens or Neuronpedia?
TransformerLens is a Python library for writing interpretability code against a model's internals. Neuronpedia is a hosted platform for exploring and sharing published features. Drowse is an interactive workbench that runs the model on your device, so you can inspect, steer, and branch without code or a server. They work well together.
Need more models?
The Python edition runs supported Hugging Face models, fits your own lenses, trains SAEs, and serves OpenAI- and Ollama-compatible APIs.
Install with pipChoose your model and tools inside the app.