Instructions to use ggml-org/Clef-Flash-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use ggml-org/Clef-Flash-GGUF 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 ggml-org/Clef-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ggml-org/Clef-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

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 ggml-org/Clef-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

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 ggml-org/Clef-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

Use Docker

docker model run hf.co/ggml-org/Clef-Flash-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use ggml-org/Clef-Flash-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ggml-org/Clef-Flash-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ggml-org/Clef-Flash-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/ggml-org/Clef-Flash-GGUF:Q4_K_M

Ollama
How to use ggml-org/Clef-Flash-GGUF with Ollama:
```
ollama run hf.co/ggml-org/Clef-Flash-GGUF:Q4_K_M
```
Unsloth Desktop

How to use ggml-org/Clef-Flash-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

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": "ggml-org/Clef-Flash-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ggml-org/Clef-Flash-GGUF with Docker Model Runner:
```
docker model run hf.co/ggml-org/Clef-Flash-GGUF:Q4_K_M
```

Lemonade

How to use ggml-org/Clef-Flash-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ggml-org/Clef-Flash-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Clef-Flash-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ggml-org/Clef-Flash-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

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 ggml-org/Clef-Flash-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ggml-org/Clef-Flash-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ggml-org/Clef-Flash-GGUF:Q4_K_M

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 "ggml-org/Clef-Flash-GGUF:Q4_K_M" \
  --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"

External validation: Clef-Flash Q4_K_M on Radeon RX 6800 (llama.cpp b11389, Vulkan, Windows) - finite typed probabilities, monotone confidence calibration on a robotics decision task

by An2333 - opened 13 days ago

Discussion

An2333

13 days ago

Setup

llama.cpp b11389, llama-server, win-vulkan-x64 build, AMD Radeon RX 6800 16 GB
Clef-Flash-Q4_K_M.gguf, flags: -ngl 99 -ub 32768 -c 16384 --host 0.0.0.0 --port 11435
(the large physical batch matches what others found: the decision head wants the whole
input covered by one batch)
Typed /v1/systemone-style decision requests; clients run inside WSL2 Ubuntu and reach
the Windows-side server over the NAT gateway — no issues
output_tokens: 0 on every single request — the no-generation, no-parse property holds

Why llama.cpp and not Ollama: Ollama 0.35.1 (Windows) fails every systemone request on
this model with Clef: non-finite logit regardless of backend (CUDA/ROCm/Vulkan/CPU,
ollama#18769). Upstream llama.cpp b11389 with this GGUF is what works — including the
smaller Q4_K_M quant, which the issue had not tested.

Evaluation (n=35 recorded decision states, robotics action-selection)

Task: at each step the model must choose the next action among 5 typed options
(observe / pick / place / verify / abort) given a canonical world-state snapshot,
for a simulated robotic workcell. Ground truth = a frozen deterministic reference
policy; we score agreement per confidence bin (offline replay, no execution):

reported confidence	n	agreements (vs reference)
[0.0, 0.5)	11	0
[0.5, 0.8)	11	9
[0.8, 1.0]	8	8

Overall agreement 17/30 on the calibration split (baseline: a generative 1.5B model
producing JSON text scored 4/30 with heavy truncation/parse failures — zero here)
The confidence bins are monotone: the model "knows when it does not know"
(low-confidence picks are 0/11), and the top bin is clean (Wilson lower bound 0.676)
Median decision latency 1.9 s, max 6.8 s (Vulkan, 16 GB VRAM, Q4_K_M)

Takeaway: Q4_K_M quantization preserves the decision head's behavior, the
probabilities are meaningfully calibrated as a routing signal (we use them to decide
when to escalate to a larger LLM), and the vulkan build makes the model practical on
consumer AMD cards under Windows.

Happy to share the evaluation harness (a small Python client for the systemone
endpoint + the binning/Wilson script) if useful.

(Evaluation harness: a small typed /v1/systemone client + Wilson binned-agreement script, available on request.)

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