Mermer123/qwen3-embedding-gguf-demo

Qwen3 Embedding GGUF Lab

An interactive, private in-browser demo for Qwen/Qwen3-Embedding-0.6B-GGUF. It runs the official Q8_0 GGUF with llama.cpp compiled to WebAssembly by wllama. No backend or GPU is required.

The Space includes:

• —semantic search over a line-separated mini corpus;
• —instructed query embeddings, following Qwen's recommended prompt format;
• —cosine similarity for a query/document pair;
• —Matryoshka dimension selection from 32 to 1024;
• —a raw embedding preview;
• —local-only browser inference with no text sent to a server.

Queries are formatted as:

Instruct: <task description>
Query:<query>

Documents are embedded without an instruction, as recommended by the model authors. Embeddings are truncated to the selected Matryoshka dimension and L2-normalized before comparison.

Browser requirements

Click Load model to download the 639 MB Q8_0 model from the Hub. It is cached by the browser when supported. A recent Chrome, Edge, or Firefox release with WebAssembly Memory64 is required; Safari is not currently supported by wllama.

The model uses last-token pooling, following the Qwen GGUF model card. All embedding computation happens on the visitor's device.