Instructions to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
# Run inference directly in the terminal:
llama cli -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
# Run inference directly in the terminal:
llama cli -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
# Run inference directly in the terminal:
./llama-cli -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

Use Docker

docker model run hf.co/tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

LM Studio
Jan

vLLM

How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-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": "tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-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/tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

Ollama
How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with Ollama:
```
ollama run hf.co/tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
```
Unsloth Desktop

How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

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": "tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with Docker Model Runner:
```
docker model run hf.co/tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS
```

Lemonade

How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

Run and chat with the model

lemonade run user.Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF-IQ3_XXS

List all available models

lemonade list

Hermes Agent

How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

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 tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS

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 "tooltd/Qwen3.8-27B-IQ4-XS-16GB-VRAM-GGUF:IQ3_XXS" \
  --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"

tooltd commited on Aug 19

Commit

7d293c2

verified ·

1 Parent(s): 43ec8f6

Update README.md

Browse files

Files changed (1) hide show

README.md +66 -0

README.md CHANGED Viewed

@@ -1,3 +1,69 @@
 ---
 license: apache-2.0
 ---

 ---
 license: apache-2.0
 ---
+# Qwen3.8-27B (GGUF — ZB-ZipBrain Quantization)
+This repository provides GGUF quantizations for **Qwen3.8-27B** optimized using **ZB-ZipBrain**, a layer-wise quantization profiling and allocation method.
+---
+## 1. Overview & Method: ZB-ZipBrain
+**ZB-ZipBrain** is an automated layer-allocation approach that dynamically profiles model layers and mixes **K-quants** and **IQ-quants** based on layer sensitivity and importance matrices.
+### Key Objectives:
+* **Selective Bit-rate Allocation:** Assigns higher precision to sensitive layers and compact IQ-quants to more resilient weights.
+* **Balanced Efficiency:** Maintains low perplexity (PPL) and minimal Kullback-Leibler (KL) Divergence relative to the BF16 baseline while achieving target file sizes / bits-per-weight (bpw).
+---
+## 2. Benchmark & Evaluation Results
+All models were evaluated against the **BF16 baseline** (`Mean PPL = 6.950493`) using standard Perplexity (PPL) and KL Divergence metrics.
+### Comprehensive Comparison Table
+| Rank | Model / File Name | Level / Source | Quant Type | Size (GB) | Mean PPL | Δ PPL | Mean KLD | Same Top-p (%) | KLD 99% |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| **1** | `Qwen3.8-27B-UD-Q8_K_XL` | unsloth | UD-Q8_K_XL | 29.30 | 6.953800 | +0.003500 | 0.000850 | 98.970% | — |
+| **2** | `Qwen3.8-27B-UD-Q6_K_XL` | unsloth | UD-Q6_K_XL | 24.14 | 6.953600 | +0.003200 | 0.001380 | 98.520% | — |
+| **3** | `Qwen3.8-27B-Q6_K` | unsloth | Q6_K | 21.31 | 6.950700 | +0.000300 | 0.002290 | 97.860% | — |
+| **4** | `Qwen3.8-27B-Q5_K_M` | unsloth | Q5_K_M | 18.47 | 6.974200 | +0.023900 | 0.006220 | 96.700% | — |
+| 5 | `Qwen3.8-27B-UD-Q4_K_XL` | unsloth | Q4_K_XL | 16.69 | 6.979220 | +0.028728 | 0.008606 | 96.091% | 0.091099 |
+| **6** | `Qwen3.8-27B-ZB4.97-GOD-IQ4_XS` | **ZB-GOD** | IQ4_XS | **15.82** | **7.004243** | **+0.053751** | **0.012249** | **95.337%** | **0.117617** |
+| **7** | `Qwen3.8-27B-ZB4.65-PRO-IQ4_XS` | **ZB-PRO** | IQ4_XS | **14.81** | **7.017278** | **+0.066786** | **0.015466** | **94.766%** | **0.150252** |
+| **8** | `Qwen3.8-27B-Q4_K_M` | unsloth | Q4_K_M | 15.93 | 6.956100 | +0.005800 | 0.015490 | 94.650% | — |
+| **9** | `Qwen3.8-27B-ZB4.60-PRO-IQ4_XS` | **ZB-PRO** | IQ4_XS | **14.65** | **7.030895** | **+0.080402** | **0.016162** | **94.668%** | **0.159115** |
+| **10** | `Qwen3.8-27B-ZB4.55-PRO-IQ4_XS` | **ZB-PRO** | IQ4_XS | **14.49** | **7.032263** | **+0.081771** | **0.016647** | **94.613%** | **0.161016** |
+| 11 | `Qwen3.8-27B-IQ4_NL` | bartowski | IQ4_NL | 15.20 | 7.006472 | +0.055980 | 0.018427 | 94.230% | 0.190168 |
+| 12 | `Qwen3.8-27B-IQ4_XS` | unsloth | IQ4_XS | 14.63 | 7.012695 | +0.062202 | 0.018652 | 94.270% | 0.194338 |
+| **13** | `Qwen3.8-27B-ZB4.48-STD-IQ4_XS` | **ZB-STD** | IQ4_XS | **14.26** | **7.050096** | **+0.099604** | **0.018892** | **94.199%** | **0.196005** |
+| 14 | `Qwen3.8-27B-Q4_K_S` | unsloth | Q4_K_S | 15.01 | 6.966826 | +0.016334 | 0.018921 | 94.235% | 0.192749 |
+| 15 | `Qwen3.8-27B-IQ4_XS-i1` | mradermacher | IQ4_XS | 14.26 | 7.012810 | +0.062318 | 0.019271 | 94.141% | 0.197891 |
+| 16 | `Qwen3.8-27B-Q4_K_S-i1` | mradermacher | Q4_K_S | 14.74 | 6.989551 | +0.039059 | 0.019805 | 93.996% | 0.204143 |
+| **17** | `Qwen3.8-27B-ZB4.36-STD-IQ4_XS` | **ZB-STD** | IQ4_XS | **13.88** | **7.054811** | **+0.104319** | **0.020556** | **93.951%** | **0.206689** |
+| 18 | `Qwen3.8-27B-Q4_0-AutoRound-Code` | webhie | Q4_0 | 14.64 | 7.067142 | +0.116650 | 0.026586 | 92.970% | 0.271619 |
+| **19** | `Qwen3.8-27B-ZB4.14-MIN-IQ4_XS` | **ZB-MIN** | IQ4_XS | **13.19** | **7.045689** | **+0.095196** | **0.029334** | **92.799%** | **0.294996** |
+| 20 | `Qwen3.8-27B-i1-IQ4_XS-GGUF-Smaller_3.96bpw` | jrell | IQ4_XS | 12.61 | 7.252766 | +0.302274 | 0.055499 | 90.090% | 0.551972 |
+---
+## 3. ZB-ZipBrain Tiers & Recommendations
+* **ZB-GOD :** God. A singularity appears. Reaches `0.012249` Mean KLD and `95.34%` top-probability match.
+* **ZB-PRO :** Pro. Recommended for 16 GB VRAM GPUs with offloading on CPU. Balances quality output with substantial size savings.
+* **ZB-STD :** Standard. Similar to other standard IQ4_XS models currently available.
+* **ZB-MIN :** Minimal. Optimal footprint for tight 16 GB memory setups, allowing headroom for longer context windows
+---
+**Note: These two missing files will be uploaded soon: Qwen3.8-27B-ZB4.97-GOD-IQ4_XS, Qwen3.8-27B-ZB4.65-PRO-IQ4_XS**
+## 4. Credits & Acknowledgements
+* **Base Model:** [Qwen3.8 27B](https://huggingface.co/Qwen) by Alibaba Cloud / Qwen Team.
+* **BF16 Base GGUF:** Provided by [Unsloth AI](https://huggingface.co/unsloth).
+* **Importance Matrix (imatrix):** Generated and curated by [ubergarm](https://huggingface.co/ubergarm).
+* **Inference & Quantization Framework:** [llama.cpp](https://github.com/ggerganov/llama.cpp) by Georgi Gerganov and contributors.