Instructions to use W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Use Docker

docker model run hf.co/W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use W4D/Qwopus3.8-27B-Coder-Flash-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "W4D/Qwopus3.8-27B-Coder-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": "W4D/Qwopus3.8-27B-Coder-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/W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Ollama
How to use W4D/Qwopus3.8-27B-Coder-Flash-GGUF with Ollama:
```
ollama run hf.co/W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M
```
Unsloth Desktop

How to use W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-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": "W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use W4D/Qwopus3.8-27B-Coder-Flash-GGUF with Docker Model Runner:
```
docker model run hf.co/W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M
```

Lemonade

How to use W4D/Qwopus3.8-27B-Coder-Flash-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Qwopus3.8-27B-Coder-Flash-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-Flash-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use W4D/Qwopus3.8-27B-Coder-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 W4D/Qwopus3.8-27B-Coder-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 "W4D/Qwopus3.8-27B-Coder-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"

Qwopus3.8-27B-Coder-Flash GGUF

GGUF quantizations of Jackrong/Qwopus3.8-27B-Coder-Flash, a Qwen3.5-architecture (27.8B, hybrid linear/full attention) coding model with vision support, converted with llama.cpp.

Highlights of this conversion:

All quants are quantized directly from the lossless BF16 conversion of the original bf16 safetensors (never requantized from a lower precision).
Quants at Q4_K_M and below use an importance matrix (imatrix) computed from the Q8_0 model over a code-heavy calibration corpus, which preserves quality in low-bit quants.
Vision support ships as a separate mmproj file (bf16) that works with every quant below.
The MTP (multi-token prediction) head from the source checkpoint is retained in every GGUF as blk.64. It is unused in normal decoding; llama.cpp can load it as a speculative-decoding drafter (the draft-mtp speculative type, implemented for this architecture in the same toolchain used for this conversion), which can provide a generation speedup. Runtime behavior with these quants has not been independently verified.

Quantizations

Sizes are measured from the produced files; bits per weight (BPW) as reported by llama-quantize.

File	Size (GiB)	BPW	Imatrix	Notes
Qwopus3.8-27B-Coder-Flash-BF16 (2 parts)	25.99 + 24.91	16.0	-	Lossless reference
Qwopus3.8-27B-Coder-Flash-Q8_0.gguf	27.05	8.50	-	Near-lossless
Qwopus3.8-27B-Coder-Flash-Q6_K.gguf	20.89	6.56	-	Very high quality
Qwopus3.8-27B-Coder-Flash-Q5_K_M.gguf	18.19	5.72	-	High quality, good balance
Qwopus3.8-27B-Coder-Flash-Q5_K_S.gguf	17.67	5.55	-	High quality
Qwopus3.8-27B-Coder-Flash-Q4_K_M.gguf	15.66	4.92	Yes	Good quality, best balance
Qwopus3.8-27B-Coder-Flash-Q4_K_S.gguf	14.74	4.63	Yes	Good quality
Qwopus3.8-27B-Coder-Flash-IQ4_XS.gguf	14.26	4.48	Yes	Good quality, smaller
Qwopus3.8-27B-Coder-Flash-Q3_K_M.gguf	12.57	3.95	Yes	Noticeable quality loss
Qwopus3.8-27B-Coder-Flash-Q3_K_S.gguf	11.41	3.59	Yes	Significant quality loss
Qwopus3.8-27B-Coder-Flash-Q2_K.gguf	10.12	3.18	Yes	Large quality loss
Qwopus3.8-27B-Coder-Flash-IQ2_M.gguf	9.54	3.00	Yes	Extreme quantization, quality loss
Qwopus3.8-27B-Coder-Flash-mmproj-BF16.gguf	0.87	16.0	-	Vision encoder (works with every quant)
Qwopus3.8-27B-Coder-Flash-imatrix.gguf	0.03	-	-	Importance matrix used for the quants (for re-quantization)

The BF16 file is split into two parts because of Hugging Face's per-file size limit. Download both parts into the same folder and point llama.cpp at the -00001-of-00002 file; it will load the second part automatically.

Downloading

# example: grab the Q4_K_M quant with the hf CLI
hf download W4D/Qwopus3.8-27B-Coder-Flash-GGUF \
  --include "Qwopus3.8-27B-Coder-Flash-Q4_K_M.gguf" --local-dir .

Usage with llama.cpp

Text-only inference:

llama-cli -m Qwopus3.8-27B-Coder-Flash-Q4_K_M.gguf -ngl 99 -c 32768

Image input (requires the mmproj file next to the quant):

llama-mtmd-cli -m Qwopus3.8-27B-Coder-Flash-Q4_K_M.gguf --mmproj Qwopus3.8-27B-Coder-Flash-mmproj-BF16.gguf --image image.png -p "Describe this image."

OpenAI-compatible server (add --mmproj for vision support):

llama-server -m Qwopus3.8-27B-Coder-Flash-Q4_K_M.gguf --mmproj Qwopus3.8-27B-Coder-Flash-mmproj-BF16.gguf -ngl 99 -c 32768

Note on context: the model uses a hybrid attention architecture (full attention every 4th layer), so the KV cache is small (~64 KiB per token) and long context is cheap.

Conversion details

Converter: convert_hf_to_gguf.py from llama.cpp master (4e910a8, 2026-11-02); quantization and imatrix tooling from llama.cpp release build b11960 (CUDA 13.4, Windows x64).
Base conversion: convert_hf_to_gguf.py --outtype bf16 --model-name Qwopus3.8-27B-Coder-Flash.
Vision projector: convert_hf_to_gguf.py --mmproj --outtype bf16.
Imatrix: computed with llama-imatrix from the Q8_0 quant, context 512, over a 2.3 MB calibration corpus of C/C++/CUDA/Python sources and documentation sampled from the llama.cpp codebase.
Quantization: llama-quantize with --imatrix for Q4_K_M, Q4_K_S, IQ4_XS, Q3_K_M, Q3_K_S, Q2_K and IQ2_M; default tensor-type policy otherwise. For IQ2_M, the 8 MTP-layer weights (blk.64.*) are pinned to Q4_K because the imatrix calibration does not cover the multi-token-prediction layer.