Explore language models with Drowse
Drowse brings language-model generation, activation steering, and mechanistic interpretability into one workbench. These guides connect the visible controls to the experiments you can run and the limits of their measurements.
Guides
- Activation steering for language models Explore activation steering without changing model weights: concepts, manifolds, lens words, SAE features, neurons and heads, ablation, and probe gates.
- Inspect tokens and internal representations Inspect token probabilities, concept probes, sparse autoencoder features, R- and S-lens layer reads, and the neurons and heads behind a token in Drowse's local LLM workbench.
- Run a language model locally in your browser Run compatible Gemma, Qwen, GPT-2, and Pythia models locally with Drowse. Learn about WebGPU, downloads, device checks, and browser storage.
- Compare language-model experiments Compare language-model continuations with Loom branches, token forks, activation interventions, template scoring, and saved experiment recipes.
Start with a model and a question
Open the browser app, run its device check, and choose a compatible model. A chat model follows a conversation template; a base model continues raw text. Generate a baseline before changing the prompt, sampling settings, or internal activations.
Choose an experiment
- Inspect a token's alternatives, its probe and feature readings, and what each layer reads and says through the S and R lenses.
- On Gemma 3 and Qwen3 models, find the neurons and heads that pushed a token, read what one writes, and push a neuron or silence a neuron or head.
- Steer with a concept, a manifold position, a lens word, an SAE feature, or a single neuron, during the whole reply or only part of it.
- Branch from a turn or token and compare continuations under controlled settings.
- Score named completion choices to inspect a distribution rather than only the sampled answer.
Find your way around
The header holds six rooms. The Workbench is where you talk to the model, with the Inspector beside the conversation. Three more read the same conversation token by token: Units (neurons and heads), Lens (each layer through both lenses) and Probes (concepts, manifolds and SAE features), and a token chosen in one is the token in the others. Artifacts is where you fit, fetch and build what they read with: packs, probes, templates and roles. Model holds the model's layers, units, device and what is prepared for it.
Extend the workbench
The hosted browser edition runs locally with WebGPU. The Python edition adds a programmatic interface, inference APIs, and instrument authoring: fitting lenses, training SAEs and fetching published unit catalogs. Visiting browser agents can use Drowse's experimental WebMCP tools when their browser supports them.