Run a language model locally in your browser
The hosted Drowse app runs compatible language models on your device using WebGPU. Prompts, generated text, activations, and saved experiments stay in browser-local storage and memory during the normal inference workflow. No Python installation or inference server is required for this edition.
Choose a chat model or a base model
The model catalog includes Gemma 3 and Qwen3 chat models, plus base-model entries such as GPT-2, Pythia 70M, Gemma 3 1B PT, and Qwen 3.5 2B Base. The app determines which entries can be installed and loaded for your device and current release. A listed model is not a promise that every GPU can run it.
Chat models use a conversation template and support model-specific role and instruction behavior. Base models continue a raw text buffer. System instructions and chat-role labels do not apply to that raw-completion workflow; use a textual scaffold or compatible activation steering.
Check WebGPU and device requirements
Drowse checks for a secure context, browser workers, hardware WebGPU, and usable local storage before loading. Browser version, GPU support, available memory, and model context size affect compatibility. The compute check is a prerequisite, not a full model-load stress test.
The browser edition requires hardware WebGPU and does not provide cloud or CPU-only inference fallback. iPhone inference remains a preview with device-specific memory constraints. If a model cannot load, follow the app's device guidance or use the Python edition on a suitable machine.
Download models and optional instruments
Models and instrument packs are downloaded from the configured distribution providers and checked against the signed catalog and artifact hashes. Downloads support recovery and progress reporting. Compatible lens and SAE packs add layer and feature inspection; availability depends on the chosen model and pack.
An installed app shell is not the same as installed model weights. Offline use depends on having the required model and runtime files stored locally. New downloads, remote artifact searches, and missing external feature descriptions need a connection.
Keep a backup of your work
Conversations autosave in this browser. Select Protect storage where available and download backups for work you want to keep. The browser can remove local data to free space, and local saves do not sync between devices.