semantic-lite GGUF

A compact, multilingual embedding model for fast local semantic search.

This repository provides production-ready GGUF releases of ukung/semantic-lite, built on a pruned Qwen3 architecture. It brings high-quality semantic understanding to local applications without requiring a large GPU or a hosted API.

Use it for semantic search, retrieval-augmented generation (RAG), similarity scoring, clustering, duplicate detection, recommendations, and multilingual knowledge bases powered by llama.cpp.

This is an embedding model, not a chat or text-generation model. Use an embedding-capable runtime and compare vectors with cosine similarity or another distance metric.

Available Files

All variants produce 1,024-dimensional embeddings and preserve the model's mean-pooling configuration.

Why semantic-lite?

• Multilingual by design: built on Qwen3's broad 100+ language coverage, including English, Indonesian, Spanish, French, German, Portuguese, Italian, Dutch, Japanese, Chinese, Korean, Arabic, Russian, Vietnamese, Thai, Turkish, Polish, Ukrainian, Swedish, Persian, Hebrew, and Hindi.
• Small and deployable: only six transformer layers, with a Q4_K_M build of about 181 MB for practical CPU and edge deployment.
• Rich semantic representations: 1,024-dimensional embeddings with mean pooling for retrieval and similarity workflows.
• Private by default: run locally with no per-request API cost and no document data leaving your infrastructure.
• Runtime-ready: works with modern llama.cpp embedding tools and servers.
• Flexible precision: choose Q4 for efficiency, Q8 for a quality-focused quantized build, or F16 as the reference release.

Benchmark Highlights

On our multilingual retrieval evaluation set, semantic-lite Q4 achieved:

The model also achieved 100% pair accuracy on the tested semantic-similarity pairs. These results demonstrate strong multilingual retrieval across English, Indonesian, Spanish, French, German, Portuguese, Italian, Japanese, Chinese, and Arabic, while keeping the model small enough for local deployment.

Languages Showcased

English · Indonesian · Spanish · French · German · Portuguese · Italian · Japanese · Chinese · Arabic

Benchmark results are measured with the Q4_K_M GGUF release on a small task-focused evaluation set. Treat them as an indicative deployment signal and validate with your own domain data.

Quick Start with llama.cpp

Build or install a recent version of llama.cpp, then run:

./llama-embedding \
	-m semantic-lite-q4.gguf \
	-p "Cara membuat nasi goreng" \
	--embd-output-format json

The command returns a JSON object containing an embedding vector with 1,024 values. Replace semantic-lite-q4.gguf with the Q8 or F16 file when you need higher numerical fidelity.

For a local HTTP service, use a recent llama-server build with embedding mode enabled:

./llama-server \
	-m semantic-lite-q4.gguf \
	--embedding \
	--host 127.0.0.1 \
	--port 8080

Then request an embedding:

curl http://127.0.0.1:8080/embeddings \
	-H "Content-Type: application/json" \
	-d '{"content":"Cara membuat nasi goreng"}'

The exact server flags and endpoint format can vary by llama.cpp version. Run ./llama-server --help for the version you installed.

Python Example

The GGUF files are intended for GGUF-compatible runtimes. One simple approach is to call llama.cpp from Python and parse its JSON output:

import json
import subprocess


def embed(text: str, model: str = "semantic-lite-q4.gguf") -> list[float]:
	result = subprocess.run(
		[
			"llama-embedding",
			"-m", model,
			"-p", text,
			"--embd-output-format", "json",
		],
		check=True,
		capture_output=True,
		text=True,
	)
	return json.loads(result.stdout)["data"][0]["embedding"]


query_vector = embed("Easy home cooking recipes")
print(len(query_vector))  # 1024

For production applications, keep the model loaded in a server process instead of launching a new process for every text.

Semantic Search

Create embeddings for documents once, normalize them, and compare a normalized query vector with a dot product. For normalized vectors, the dot product equals cosine similarity:

import numpy as np

documents = [
	"How to cook fried rice",
	"Guide to stock market investing",
	"Tips for caring for your cat",
]

document_vectors = np.array([embed(document) for document in documents])
document_vectors /= np.linalg.norm(document_vectors, axis=1, keepdims=True)

query = np.array(embed("Easy home cooking recipes"))
query /= np.linalg.norm(query)

scores = document_vectors @ query
best_index = int(np.argmax(scores))
print(documents[best_index])

Choosing a Variant

• Choose Q4_K_M for the lowest memory use and fast local CPU inference.
• Choose Q8_0 when retrieval quality is more important and additional memory is available.
• Choose F16 as the high-fidelity reference for evaluation or quality-sensitive workloads.

Quantization can slightly change similarity scores. When comparing benchmark results, use the same GGUF variant consistently for both indexing and querying.

Limitations and Compatibility

• This repository contains GGUF exports of the upstream model, not a new training checkpoint.
• Use the same preprocessing, normalization, and pooling behavior for documents and queries.
• The model is intended for embeddings, not text generation.
• Runtime support depends on using a recent llama.cpp version with Qwen3 embedding support.
• The upstream model card documents multilingual coverage and recommended retrieval prompts such as query: and passage: where appropriate.

License

The upstream model is released under the Apache-2.0 license. See the upstream model card for the original model details and license information.