runonweb.ai

Free, open-source AI modules that run in the browser. Live demos + the runonweb package.

• One import per task: runonweb/stt, runonweb/remove-bg, runonweb/embed…
• WebGPU when available, WASM fallback. Nothing leaves the device.
• Every module is built on an open model. Each demo page credits the base model, author and license. The code is MIT; the models keep their own licenses (see License).

pnpm install
pnpm dev        # astro dev --background is preferred, see AGENTS.md
pnpm check      # astro check (types)
pnpm build

Monorepo: the site lives in packages/web (Astro + Cloudflare Workers) and the library in packages/runonweb. Root scripts forward to web. Every push to main is built and deployed by Cloudflare Workers Builds on the runonweb Worker (root directory /packages/web, pnpm install --frozen-lockfile && pnpm run build, then npx wrangler deploy; build variables PNPM_VERSION=11.28.0 and SKIP_DEPENDENCY_INSTALL=1).

Package

import { SpeechToText } from 'runonweb/stt'
import { RemoveBackground } from 'runonweb/remove-bg'
import { ImageCaptioner } from 'runonweb/caption'
import { DepthEstimator } from 'runonweb/depth'
import { ObjectDetector } from 'runonweb/detect'
import { TextEmbedder, cosineSimilarity } from 'runonweb/embed'
import { Classifier } from 'runonweb/classify'
import { Translator, LANGS } from 'runonweb/translate'
import { Emojifier } from 'runonweb/emoji'
import { TextToSpeech } from 'runonweb/tts'
import { isWebGPUAvailable } from 'runonweb/core'

Source and API docs: packages/runonweb.

Modules

Module	Base model	License	Download (WebGPU / WASM)
runonweb/stt	Whisper tiny.en (OpenAI)	Apache-2.0	~150 MB / ~40 MB
runonweb/remove-bg	BEN2	MIT	WASM only · ~219 MB
runonweb/caption	LFM2.5-VL-450M (Liquid AI) · 9 languages	LFM Open License v1.0 (free < USD 10M revenue)	~316 MB / ~505 MB
runonweb/depth	Depth Anything V2 Small	Apache-2.0	~50 MB / ~27 MB
runonweb/detect	RF-DETR Nano (Roboflow)	Apache-2.0	WASM only · ~29 MB
runonweb/ocr	PP-OCRv6 tiny / small / medium (PaddlePaddle)	Apache-2.0	~6 MB / ~31 MB / ~139 MB (same files on both)
runonweb/embed	all-MiniLM-L6-v2 (sentence-transformers)	Apache-2.0	~45 MB / ~23 MB
runonweb/classify	Kev-0.8B (Jared Palmer) on Qwen3.5-0.8B-Base, LoRA merged + ONNX by runonweb (training/kev-onnx)	Apache-2.0	~750 MB · large (Kev-4B) ~2.7 GB (same files on both)
runonweb/translate	Firefox Translations (Mozilla, Marian NMT via Bergamot WASM) · optional M2M100 418M (Meta)	MPL-2.0 · MIT (M2M100)	WASM only · ~21–55 MB per pair · M2M100 ~640 MB
runonweb/clean	S1-mini (Superwhisper)	Apache-2.0 + naming term	~365 MB / ~413 MB
runonweb/emoji	text2emoji-tiny (trained by runonweb, training/text2emoji)	MIT	WASM only · ~4 MB
runonweb/tts	Kokoro 82M (hexgrad) / Supertonic 2 (Supertone) / KittenTTS nano	Apache-2.0 · OpenRAIL-M (Supertonic)	~326 MB (WebGPU) or ~92 MB (WASM) / ~262 MB / ~28 MB
runonweb/image	Bonsai Image 4B binary / ternary (Prism ML, on FLUX.2 Klein)	Apache-2.0	WebGPU only · ~3.4 GB / ~3.9 GB

Sizes are approximate. The catalog (names, colors, base model, license) lives in packages/web/src/data/models.ts.

What we learned testing backends (Transformers.js 4.2, ONNX Runtime Web)

• q8 on WebGPU silently produces garbage (embeddings had ~0.6 similarity for unrelated text). Use fp16/fp32 on WebGPU, q8 on WASM.
• Encoder-decoder models (Whisper aside) fail at session creation on WebGPU: ViT-GPT2, OPUS-MT, M2M100. They are pinned to WASM via supportedDevices: ['wasm']. Kokoro TTS runs on WebGPU.
• Translation left Transformers.js: Mozilla's Firefox Translations models (Bergamot WASM, int8) are 3–5× smaller than OPUS-MT q8 at equal or better COMET, and cover 106 pairs. Weights are fetched with pnpm models:translate into weights/firefox-translations/ (gitignored).
• ORMBG and BiRefNet fail in ONNX Runtime Web (ceil() MaxPool unsupported / abort). MODNet works but is portrait-oriented.
• BEN2 (the remove-bg model) loads on WebGPU, then OrtRun fails: its fused LayerNorm is fp16 in / fp32 scale+bias+out, and the shader does not compile (Invalid ShaderModule "LayerNorm"). Pinned to WASM until onnxruntime-web includes onnxruntime#32629.
• On a laptop GPU, WebGPU fp16 is 5–7× faster than WASM q8 for vision models.
• The default onnxruntime-web entry is the JSEP build: its MatMulNBits only runs 2/4-bit weights, so 8-bit layers fail at session creation (nbits_ == 4 || nbits_ == 2 was false). onnxruntime-web/webgpu (the native WebGPU EP, also what Transformers.js uses) runs them, but its CPU side has no GatherBlockQuantized (int4 embeddings), so the WASM fallback must load onnxruntime-web/wasm. runonweb/classify picks the entry per backend.

Adding a module

1. Create packages/runonweb/src/<name>/index.ts using loadPipeline from core/pipeline.ts. Pick an explicit dtype per device and comment the resulting download size.
2. Add the export to packages/runonweb/package.json.
3. Add a Model entry to packages/web/src/data/models.ts with name, color, base model, author, license and size.
4. Add packages/web/src/pages/models/<slug>.astro + packages/web/src/scripts/models/<slug>.ts. Reuse ModelShell, DemoStatus, DemoButton, ImageDropzone and createDemoUI.
5. Test the model in the browser on both backends before shipping; see the backend notes above.
6. Only use models whose license allows redistribution and commercial use (Apache-2.0, MIT, BSD, CC-BY). Avoid -NC licenses.

License

The code in this repository and the runonweb package is MIT.

The MIT license does not cover the models. The package ships no weights: the browser downloads them from the Hugging Face Hub on first use, and each model keeps the license its authors chose. The Modules table lists them per module, and MODELS.md per variant, with links to the three that add terms of their own:

• runonweb/caption: LFM Open License v1.0, free for individuals and companies under USD 10M annual revenue.
• runonweb/tts with size: 'multi': OpenRAIL-M, with use restrictions (no impersonation, deception or illegal use).
• runonweb/clean: Apache-2.0, plus the duty to keep calling the model "S1-mini" by "Superwhisper".

The fonts in packages/web/src/assets/fonts are Geist, under the SIL Open Font License 1.1 (LICENSE).