# FreeGen

> Free browser tools that run open AI models entirely on your own computer through WebGPU. No account, no upload, no cost, no server.

FreeGen runs chat, speech to text, text to speech, image understanding, image editing, and image generation locally in the browser. Model weights are downloaded once from Hugging Face and cached; prompts, audio, and images never leave the device. Every model offered is under a licence permitting commercial use, and the licence is shown before download. Requires a browser with WebGPU on a laptop or desktop.

## Studios

- [Chat](https://freegen.ai/chat): A local assistant for writing, code, and reasoning. Streams tokens straight from your GPU. Nothing you type is sent anywhere.
  - SmolLM2 360M (Apache 2.0, 260 MB, Instant). Quickest to download. Good for short drafting and rewriting.
  - Qwen3 0.6B (Apache 2.0, 543 MB, Balanced). The default. Noticeably better reasoning for a modest download.
  - Qwen3 1.7B (Apache 2.0, 1.3 GB, Capable). Strongest answers. Needs a roomy GPU and patience on first load.
- [Transcribe](https://freegen.ai/transcribe): Record or drop in audio and get an accurate transcript. Whisper runs on your machine, so private recordings stay private.
  - Whisper Tiny (Apache 2.0, 114 MB, English only). Fastest option. English audio only.
  - Whisper Base (Apache 2.0, 197 MB, Multilingual). Handles 90+ languages and can translate to English.
  - Whisper Large v3 Turbo (MIT, 538 MB, Most accurate). The strongest transcription available in a browser. Slower to load, noticeably better on accents and noise.
- [Speak](https://freegen.ai/speak): Read anything aloud in a natural voice and save the audio. A compact neural voice with no per-character billing.
  - Kokoro 82M (Apache 2.0, 86.0 MB, Standard). 28 voices. Runs on the CPU, so it works without WebGPU.
  - Kokoro 82M full (Apache 2.0, 310 MB, Best quality). The same voices at full precision. Cleaner, but a much bigger download.
  - Supertonic (OpenRAIL, 250 MB, Fastest). Ten voices at 44.1 kHz, the highest quality here, and quick to warm up.
- [Look](https://freegen.ai/look): Ask questions about a picture, or find every object in it. Captions, object detection, text extraction, and free-form questions.
  - Describe it (MIT, 221 MB). Writes a few sentences about the picture. Good for alt text.
  - Find objects (MIT, 221 MB). Boxes and labels everything it recognises.
  - Read the text (MIT, 221 MB). Pulls out any writing in the image.
  - Ask a question (Apache 2.0, 635 MB). Ask anything about the picture in your own words. A bigger download than the fixed jobs above.
- [Edit](https://freegen.ai/edit): Click to cut out any object, upscale 2x, or build a depth map. Full-resolution results that never leave the browser.
  - Remove background (Apache 2.0, 12.0 MB). Best on people and single subjects. Exports a transparent PNG.
  - Remove an object (Apache 2.0, 228 MB). Click something to select it, then remove it and let the background be filled in.
  - Upscale 2x (Apache 2.0, 31.0 MB). Doubles resolution. Keep the input under about 1000px.
  - Click to select (Apache 2.0, 20.0 MB). Click any object to cut just that out. Shift-click to remove a region.
  - Depth map (Apache 2.0, 18.0 MB). Greyscale distance map for compositing and masking.
- [Shape](https://freegen.ai/shape): Turn one photograph into a model you can open in Blender. Reconstruction runs on your GPU, so the photo never leaves the machine.
  - TripoSR (MIT, 485 MB, Single photo). About 485 MB on first use, then cached. A model takes roughly a minute. Photographs get a 12 MB background remover as well, because the reconstruction expects a cut out subject.
  - TripoSG (MIT, 3.4 GB, Best quality). A much better shape than TripoSR, for a much larger download and a couple of minutes of work. Geometry only, with no colour.
- [Imagine](https://freegen.ai/imagine): Describe a picture and watch your own GPU draw it. The heaviest thing here by a wide margin, and the most fun.
  - Janus-Pro 1B (MIT, 1.5 GB, 384px). About 1.5 GB on first use, then cached. Expect roughly a minute per image.
  - FLUX.2 Klein 4B (Apache 2.0, 6.4 GB, 512px). The best quality here by a wide margin, and the largest download by a wide margin too.

## Other pages

- [Storage](https://freegen.ai/data): See and manage every model you have downloaded and everything you have made, all stored in this browser.
- [Library](https://freegen.ai/library): Browser tools from across the web, ranked against what your machine can actually run.
- [Latest](https://freegen.ai/latest): Runtime releases and newly published browser tools, straight from the official feeds.

## Notes

- Everything is free. There is no paid tier, no trial, and no account.
- Nothing is uploaded. Inference happens on the visitor's own graphics card.
- The image model is FLUX.2 Klein 4B, exported to ONNX by us from the Apache-2.0 upstream and published at https://huggingface.co/cgb/flux2-klein-4b-onnx-webgpu
- Source of the weights and export notes: https://huggingface.co/cgb/flux2-klein-4b-onnx-webgpu