Instructions to use thinletter/qwen3-embedding-0.6b-query-clients with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use thinletter/qwen3-embedding-0.6b-query-clients with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
# Run inference directly in the terminal:
llama cli -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
# Run inference directly in the terminal:
llama cli -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
# Run inference directly in the terminal:
./llama-cli -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Use Docker

docker model run hf.co/thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

LM Studio
Jan
Ollama
How to use thinletter/qwen3-embedding-0.6b-query-clients with Ollama:
```
ollama run hf.co/thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
```
Unsloth Desktop

How to use thinletter/qwen3-embedding-0.6b-query-clients with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "thinletter/qwen3-embedding-0.6b-query-clients:Q4_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use thinletter/qwen3-embedding-0.6b-query-clients with Docker Model Runner:
```
docker model run hf.co/thinletter/qwen3-embedding-0.6b-query-clients:Q4_0
```

Lemonade

How to use thinletter/qwen3-embedding-0.6b-query-clients with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Run and chat with the model

lemonade run user.qwen3-embedding-0.6b-query-clients-Q4_0

List all available models

lemonade list

Hermes Agent

How to use thinletter/qwen3-embedding-0.6b-query-clients with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use thinletter/qwen3-embedding-0.6b-query-clients with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf thinletter/qwen3-embedding-0.6b-query-clients:Q4_0

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "thinletter/qwen3-embedding-0.6b-query-clients:Q4_0" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Query-side clients for Qwen3-Embedding-0.6B: keep your index, shrink the query encoder — summary and links

by honza-rosecky - opened 1 day ago

Discussion

honza-rosecky

thinletter org 1 day ago

Four GGUF query encoders (340 / 385 MiB instead of 1 142 MiB fp16) that stay compatible with document vectors produced by the unchanged Qwen3-Embedding-0.6B: the index is not touched, only the query side moves to the client. On the model's own fp32 indices the Q4_K_M file keeps 100.0 / 99.4 / 100.1 / 99.3 % of nDCG@10 on SciFact / NFCorpus / ArguAna / SciDocs (cosine to the fp32 query vector 0.97–0.98); the file calibrated on Czech text keeps 98.7 % on a 55 071-segment Czech supreme-court index. Q5_K_M (385 MiB) sits 0.01 closer in cosine and is otherwise within noise.

Along the way: every ≤ 3-bit file of this model failed our release rule (nDCG holds on English while the query vectors drift to cosine 0.90–0.94; on Czech nDCG falls to 80–93 %), and a 2-bit token table alone costs 4.6–6.6 points on Czech. The same architecture as a retrieval fine-tune (harrier-0.6b) holds 99 % at 3.4 bits, so the tolerance is a property of the checkpoint, not of the recipe. The official sentence-transformers prompt has no trailing space after Query:; the files were evaluated with that prompt.

Pre-registered runs, paired bootstrap over queries, per-checkpoint compatibility stated on the card:

verify against your own index (recipe + harness): https://github.com/rosecky/embedding-quantization-public
report: https://github.com/rosecky/embedding-quantization-public/blob/main/docs/release/technical_report.md
browser demo (wllama, WebGPU/WASM): https://thinletter.io/demo

Limits: one Czech corpus with synthetic queries and no human judgements; differences under 0.01 nDCG@10 are ties (calibration-draw variance). Questions and attacks welcome here.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment