Instructions to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head 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 robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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 robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
# Run inference directly in the terminal:
./llama-cli -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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 robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS

Use Docker

docker model run hf.co/robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS

LM Studio
Jan
Ollama
How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with Ollama:
```
ollama run hf.co/robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
```
Unsloth Desktop

How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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": "robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with Docker Model Runner:
```
docker model run hf.co/robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS
```

Lemonade

How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS

Run and chat with the model

lemonade run user.qwen3.8-27b-ud-iq2xxs-q8head-UD-IQ2_XXS

List all available models

lemonade list

Hermes Agent

How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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 robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_XXS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use robbintt/qwen3.8-27b-ud-iq2xxs-q8head with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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 "robbintt/qwen3.8-27b-ud-iq2xxs-q8head:UD-IQ2_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"

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Qwen3.8-27B UD-IQ2_XXS + Q8_0 LM Head (graft)

Custom GGUF for Qwen3.8-27B built by grafting the Q8_0 output/lm_head from the Q8_0-tier Unsloth Dynamic quant onto the IQ2_XXS (Unsloth Dynamic) body.

Base: unsloth/Qwen3.8-27B-GGUF -> Qwen3.8-27B-UD-IQ2_XXS.gguf
Body: IQ2_XXS (unchanged, byte-identical to base)
LM head (output.weight): Q8_0 (1,288 MiB), up from the base Q3_K (521 MiB)
output_norm: F32 (unchanged)
quantization_version: 4
Size: 7.52 GiB (base 6.77 GiB, +767 MiB)

Why: the lm_head is only ~5-9% of the file but is the single highest-leverage tensor for output quality. Holding it at Q8_0 on a 2-bit body yields better next-token logits for a small size cost.

How: header-offset GGUF surgery — output.weight (tensor #1, offset 0) replaced with a Q8_0 blob; all 850 following tensor offsets shifted +767 MiB. Body bytes are byte-identical to the base.

GGUF alignment note (the bug that mattered): GGUF tensor data does not begin at header_end — it begins at the 32-byte-aligned boundary align_up(header_end, 32) (llama.cpp reads it at GGUF_PAD(header_end, 32)). The Q8_0 head was sourced from Q6_K[align_up(Q6_header_end,32)] and the body from base[align_up(base_header_end,32) + old_head], with 3 bytes of alignment padding written after the header in the output.

Validated: loads clean in llama.cpp and produces coherent output:

"What is the capital of France?" -> "The capital of France is Paris."
"What is 2+2?" -> "4"
Fluent paragraph generation on a 2-bit body with the Q8_0 head (~43 t/s on a single RTX 3090).
Dequant of the grafted output.weight row-0 matches the official BF16 lm_head (corr ~1.0, max abs err ~3.6e-4).

Earlier revisions of this repo (before the alignment fix) produced garbage ("/////////") because the head was copied from the un-aligned header_end. The current file is the corrected build.

Downloads last month: 604

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

2-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support