Model & settings
Guide — run an AI model in your browser
The strip above the message box shows the model picked for your browser: a GPU model when WebGPU works, a small CPU model when it does not. One press of Start loads it.
You see percent, megabytes, speed and time left, and you can cancel at any time. The files stay in your browser, so the next start takes seconds.
Type a message, or drop a PDF, text file or image anywhere on the page. Files are read on your device; images need the Phi 3.5 Vision model.
The Tune tab of the side panel (Alt+,) has personas, temperature, top-p and reply length, or type /temp 0.2 in the message box. Think shows step-by-step reasoning on models that support it.
Chats are saved in this browser: search, rename, delete with undo, export as Markdown or JSON. Focus mode (Alt+X) hides everything but the chat; Ctrl+K opens every command.
Which model for which job
- Qwen 3 0.6B
- 335 MB, the smallest GPU model with a Think switch
- Llama 3.2 1B
- 695 MB, a dependable default for everyday chat and rewrites
- SmolLM2 135M
- 137 MB on the CPU, for phones and browsers without WebGPU
- Qwen 3 4B
- 2.26 GB, the usual pick for a desktop GPU, Think on or off
- Phi 4 Mini
- 2.16 GB, a compact model from Microsoft
- Gemma 4 E2B
- ~2.5 GB (an estimate), a small model from Google
- Qwen 3
- a Think switch: reason step by step, or answer directly
- DeepSeek R1 Distill
- 1.5B, 7B and 8B builds that reason before every answer
- Ministral 3 Reasoning
- 1.93 GB, the reasoning edition of the 3B model
- Phi 3.5 Vision
- 2.77 GB, the one model here that reads images
- Qwen 2.5 Coder
- 1.5B, 3B and 7B builds tuned for programming
- Qwen 3.5 9B
- 5.04 GB, the largest download here; needs a powerful GPU
Free browser-based AI, explained
A browser-based LLM is a language model that runs inside a web page instead of on a company’s server. This page downloads a model once, keeps it in your browser’s storage and does every calculation on your own graphics card or processor. There is no account, no API key and no usage limit other than what your hardware can do.
The workspace
On a wide screen the page has three panes: your saved chats on the left, the conversation in the middle and a model and settings panel on the right, with Model, Tune and More tabs. A status line underneath shows the engine, the speed of the last reply in tokens per second and how full the context window is. The model name at the top switches with one click between the models already on this device, and an empty chat offers prompt starters that fill in the message box for you. On a phone the two side panes open as drawers.
Two engines: GPU and CPU
The 31 GPU models run with WebLLM 0.2.86 on WebGPU. They are 4-bit builds (q4f16_1) from the MLC project and are the fast option. The 6 CPU models run with transformers.js 4.3.0 inside a Web Worker on WebAssembly, using 8-bit weights, which run noticeably faster on a CPU than 4-bit ones. They are slower than GPU models, but they work on phones, older browsers and computers without a usable GPU, and the page stays responsive while they answer.
What your device needs
WebGPU ships in Chrome and Edge 113 and newer on desktop, Chrome 121 and newer on Android, Firefox 141 and newer on Windows, and Safari 26. The GPU models in this list also need the shader-f16 GPU feature. The device check in the Model tab of the side panel asks your browser for both and tells you plainly when one is missing, then recommends a CPU model instead. The GPU memory column below is the figure WebLLM lists for each model; the page compares it with the limits your GPU reports.
Models at a glance
| Model | Engine | Download | GPU memory | Note |
|---|---|---|---|---|
| SmolLM2 135M | CPU | 137 MB | — | Smallest download, for phones |
| SmolLM2 360M | CPU | 365 MB | — | Default where WebGPU is missing |
| Qwen 3 0.6B | GPU | 335 MB | 1.4 GB | Smallest GPU model with a Think switch |
| Llama 3.2 1B | GPU | 695 MB | 0.9 GB | Default on smaller GPUs |
| Qwen 3 1.7B | GPU | 968 MB | 2.0 GB | Think switch |
| Llama 3.2 3B | GPU | 1.81 GB | 2.3 GB | General chat |
| Qwen 3 4B | GPU | 2.26 GB | 3.4 GB | Default on larger GPUs, Think switch |
| Phi 4 Mini | GPU | 2.16 GB | 3.4 GB | Compact model from Microsoft |
| Gemma 4 E2B | GPU | ~2.5 GB | 3.4 GB | Download size is an estimate |
| Phi 3.5 Vision | GPU | 2.77 GB | 4.0 GB | Reads images |
| Qwen 2.5 Coder 7B | GPU | 4.28 GB | 5.1 GB | Tuned for code |
| DeepSeek R1 Distill 7B | GPU | 4.28 GB | 5.1 GB | Reasons before it answers |
| Mistral 7B v0.3 | GPU | 4.08 GB | 4.6 GB | General chat |
| Llama 3.1 8B | GPU | 4.52 GB | 5.0 GB | General chat |
| Qwen 3 8B | GPU | 4.61 GB | 5.7 GB | Think switch |
| Qwen 3.5 9B | GPU | 5.04 GB | 6.4 GB | Largest download in the list |
The picker lists all 37 models with search and filters (fits my device, on this device, under 1 GB, reasoning, vision, code, multilingual, CPU). Download sizes are the page’s catalogue values; the browser keeps each model after its first download.
Chat with PDFs and text files
Attach a PDF, text, Markdown, CSV, JSON or code file with the paperclip, or drop it anywhere on the page. PDFs are read with pdf.js, which extracts the text layer only. A scanned PDF has no text layer and this tool does not do OCR, so it tells you that instead of guessing. Short files go to the model whole. Long files are cut into overlapping chunks, embedded on your CPU with IBM Granite Embedding (97M, multilingual), and only the four best-matching chunks are added to your question. You can attach up to 5 files of 25 MB each. Attached files are used in the current session and are not saved with the conversation: only their names are.
Images
Phi 3.5 Vision is the model that reads images. Attach one or paste a screenshot; the page shrinks it to at most 1,024 pixels on its longest side before the model sees it. If the loaded model cannot read images, the page says so and offers to switch to the vision model.
Reasoning, JSON mode and personas
Qwen 3 models have a Think switch. DeepSeek R1 distills and Ministral 3 Reasoning always reason first. The reasoning is shown in a collapsed block labelled “Thought for N s”, so the answer stays easy to read. JSON mode (GPU models) constrains the reply to valid JSON. Personas are ready-made system prompts: assistant, coder, translator, summariser and strict JSON; you can edit the text freely. Temperature, top-p (GPU models) and the maximum reply length (up to 8,192 tokens on GPU models, 2,048 on CPU models) sit in the Tune tab of the side panel. Qwen 3 supports 119 languages and dialects according to its developers, which makes it a sensible first try for Chinese, Dutch or other non-English text.
Context window
GPU models run with a 4,096-token context window here. When a conversation outgrows it, the oldest turns are left out of what is sent to the model (they stay on screen) and the meter in the status line under the chat shows how full the window is, in tokens.
Conversations, export and shortcuts
Conversations are stored in your browser’s localStorage (up to 80). Search them by title or text, rename or delete them (delete has an undo), export one as Markdown or JSON, export everything, or import a JSON backup. Keys: Enter sends, Shift+Enter adds a line, Esc stops a reply, ↑ edits your last message, Alt+N starts a new chat, Alt+M opens the models, Alt+L toggles the chat list, Alt+, shows or hides the model and settings panel, Alt+X toggles focus mode, Alt+R regenerates, Alt+C copies the last answer, Alt+/ shows the list, Ctrl+K (⌘+K on a Mac) opens the command palette, and typing / at the start of the message box lists slash commands such as /model, /system, /temp, /think and /export. Add ?lang=zh to the address for the Chinese interface or ?mode=focus to open in focus mode.
Privacy: what stays and what leaves
This page has no backend. Its own code never sends your messages, files or images anywhere; prompts and replies are processed in memory in your tab. The page does use the network for things that are not your content: the JavaScript libraries (WebLLM from esm.run, transformers.js and pdf.js from jsDelivr), the model files (weights from Hugging Face, the compiled WebGPU model library from the WebLLM project’s GitHub repository, and the embedding model the first time you attach a long document), the JetBrains Mono font from Google Fonts, and SimpleAnalytics, a cookie-free page-view counter used on every page of this site. Saved chats stay in localStorage; the “Save chats in this browser” switch turns saving off. Clearing site data removes them.
Honest limits
These are small models. They are useful for drafts, rewrites, summaries, short code and extracting data, and they get facts wrong, so check anything important. They are slower than cloud chatbots, and how much slower depends on your GPU. Large models need a lot of GPU memory; if the browser loses the GPU, reload the page. The page is built as a demonstration and has not been tested for heavy use.