Semantic search is now open, fast, free and available for everyone
in your browser & Node.js for fast search < 100 ms for ~97% recall @ top 5 with just 39.8 MB model all-in and ~21,5 kB index for 100 docs who need fast, efficient semantic search
The Winzling embedding model is an extremely small text embedding model created with specific quantization and pruning methods. With a model size of 39.8 MB all-in it achieves ~97% recall in German, English and other languages at top 5 - usually within < 100ms even on the weakest devices and in-browser.
Model, inference and vector search algorithms brought to you by Aron Homberg (kyr0) (opens in a new tab) – Open Source (MIT (opens in a new tab))
Small enough to ship
4-bit TurboQuant stores each document in 256 bytes, so a whole index travels to the client like an image does. Vectors have 384 dimensions; a text may run to 8,192 tokens.
Runs on modest hardware
One codebase for Node.js and the browser, on CPU/WASM with no GPU: < 100 ms end to end and about 216 MB RAM. Made for old phones, weak business laptops and embedded boards.
Open Research, Customizable
make bench regenerates every figure on this page from a tiny benchmark (opens in a new tab). The method carries over to other languages, multimodal, late-interaction and fine-tuned embedders. Do you need a custom model? Ask Aron Homberg (LinkedIn) (opens in a new tab).
Zero cost - it's free!
MIT-licensed code, commercially usable, public model files on Hugging Face and static hosting: no API keys, no per-query fees, no servers to run. This page is plain files.
Perfect when privacy matters
Queries and your own documents never leave the device: no server, no third-party AI API, no data transfer to justify under GDPR (DSGVO). Only the model and runtime files are fetched; self-host them to keep even that in-house.
Pareto-optimal quality
Good enough for most retrieval and ranking: the right passage is in the top 5 for ~97% of German, English and Russian queries (92.2% across all nine target languages) and in the top 25 for 98.9%, at 39.8 MB instead of 305 MB for the larger Harrier. Measured on 2,000 passages.
Tokenize.
Embed.
Retrieve.
🔍 Finding a needle in a haystack
100 Belebele passages, each in 20 languages: 2,000 passages in one index. Pick one of the 1,000 example queries with ⌘K, or type your own. Or create your own vector database index right in your browser, in the tab.
Runtime
Loading the index. The model loads with your first search.
Backend
–
ONNX Runtime Web, 1 thread
Ready in
–
download, compile, warm-up
Vector Database Index
–
documents
Last query
–
embed + TurboQuant scan
for
Clear searchSearch 2,000 passages in 20 languages
Enter a question above and press Search, or pick one of the 1,000 example queries. The first search downloads the 39.8 MB model once; every search after that runs on this device as you type.
Your notes
Write anything, in any language. A note is embedded 500 ms after you stop typing, or when you leave it; an empty note leaves the index.
No notes yet
Add a note with the button above. Each one becomes a 384-dimensional vector in your browser.
for
Clear searchSearch what you wrote
Add a few notes on the left, then ask a question here. Matching is by meaning, across languages, not by shared words.
💡 One query, three steps, all on your device
Documents are embedded once, ahead of search: this demo's passages in Node.js, your own notes right in the browser. A search embeds only the query and compares it with the compressed vectors.
- 39.8 MB, cached
- text
- vectors
- codes
- 431 kB · once
- text
- 384-d
- top k
Measured, not promised*
Winzling in headless Chromium (WASM, 1 thread) on the same 2,000-passage haystack, from make bench. Winzling-Embed-a8m-64k (opens in a new tab) is a pruned model that is further pruned for a few select languages, and then custom quantized to extremely shrink it in size. Therefore, low performance in non-target languages is by design.
* measured on an Apple MacBook Air M4 in a Chromium-based browser.
R@5 by query language
Share of a language's 50 questions whose gold passage ranks in the top 5 of 2,000. Chromium, TurboQuant. Target: the nine languages Winzling is built for. Pruned: below 70% R@5.
Need it tuned to your languages and data?
Aron Homberg, freelance and independent AI researcher: custom embedding models, vector search and on-device AI, offers consulting services worldwide.
🧑💻 Add it to your own app
Write the six lines yourself, or hand one prompt to your coding agent.
NPM package
6 lines in your frontend app. The first call downloads the pinned model from Hugging Face, checks every file's SHA-256 and caches it in the browser.
import { buildTurboQuantIndex, createWinzlingEmbedder, searchTurboQuantIndex } from "defuss-vectorsearch/browser.js";const embedder = createWinzlingEmbedder();const documents = ["The cat sleeps on the sofa.", "Die Börse schloss heute im Plus.", "Кошка спит на диване."];const index = buildTurboQuantIndex(await embedder.embedDocuments(documents));const hits = searchTurboQuantIndex(index, await embedder.embedQuery("A sleeping cat on a couch."), 2);console.log(hits.map((hit) => documents[hit.index])); // both cat sentencesVibe coding
Paste this prompt into your coding agent. It links the repository, so the agent works from the current README.
Integrate defuss-vectorsearch into this app: https://github.com/kyr0/defuss-vectorsearchRead its README first and follow the browser quick start: import from "defuss-vectorsearch/browser.js", embed with createWinzlingEmbedder() inside a Web Worker, index with buildTurboQuantIndex() and search with searchTurboQuantIndex().Embed the documents once ahead of time, in Node.js or in the browser, and keep the index as a static file or in memory; a search then embeds only the query.The first use downloads the pinned Winzling model (39.8 MB) and caches it in the browser, so start it when the user first searches.⏎ sendEmbed, index and search anywhere, for free
Running an embedding model, building a vector index and searching it now works wherever Node.js or a browser runs: no GPU, no server, no per-query API bill.
MIT licensed · Winzling model and Belebele data credited below