Instructions to use empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M

Use Docker

docker model run hf.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "empero-ai/Qwen3.8-35B-A3B-Distill-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": "empero-ai/Qwen3.8-35B-A3B-Distill-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M

Ollama
How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with Ollama:
```
ollama run hf.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
```
Unsloth Desktop

How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf empero-ai/Qwen3.8-35B-A3B-Distill-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": "empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with Docker Model Runner:
```
docker model run hf.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M
```

Lemonade

How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.8-35B-A3B-Distill-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-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 empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use empero-ai/Qwen3.8-35B-A3B-Distill-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf empero-ai/Qwen3.8-35B-A3B-Distill-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 "empero-ai/Qwen3.8-35B-A3B-Distill-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"

Ternary-based 35B-A3B I made, using this distill.

by SkyIsNotGreen - opened 9 days ago

Discussion

SkyIsNotGreen

9 days ago

•

edited 9 days ago

Thanks for the distill. I quantized Qwen3.8-35B-A3B-Distill down to 2.61 bpw and wanted to show you the result.

Scion-35B-A3B: the expert banks in a ternary container (2-bit codes plus one fp16 group scale per 128 weights), everything else Q8_0, plus small trained rank-512 corrections (attention output, MoE block output and router deltas), trained by output-KD against the BF16 teacher with the deployed quantizer in the loop. No imatrix, no calibration corpus, no full-model QAT. The corrections are the only trained part, and the run cost about $7 of rented H100 time.

11.34 GB / 2.61 bpw, single file (corrections embedded, no --lora), 6.3x smaller than the BF16 reference.
Task retention (400 tasks each): HellaSwag 79.00 and Winogrande 76.25 against BF16's 81.25 and 76.00, inside the ±2% noise band; best PPL of the 2-bit class (8.354 versus IQ2_M's 8.413).
The gap: full-vocabulary KLD against BF16 is still 2-bit-class (0.269 mean versus Q4_K_M's 0.031). Tail-aware training is my next step.

Apache-2.0, inherited, with attribution to empero-ai and Qwen (and the community BF16 GGUF conversion). It needs my llama.cpp fork (PQ2_0 plus embedded adapters; stock llama.cpp cannot load the tensor types). I built it to serve a local assistant stack on a single 20 GB card.

Card and weights: https://huggingface.co/SkyIsNotGreen/Scion-35B-A3B
Build-scripts, write-up and failed routes: https://github.com/sky-is-green/scion

If anyone wants to try it or poke holes in the numbers, the model repo's discussions are open, and I am happy to share any part of the recipe in more detail.

MultivariatePerspective777

9 days ago

Do you plan to use SignRoundV2 and/or CAT-Q? https://arxiv.org/html/2512.04746v2 https://arxiv.org/html/2606.26650v1

I suspect this Ternary training harness might be useful:
https://huggingface.co/penkia/TernaryQuench-Qwen3.8-27B-GGUF
They did something similar with 27B, but I suspect if SignRoundV2 is integrated the results would improve.

That said, did you use these, and if not what else did you use?

I tell you out of my own self-interest, I'd like the 35B-A3B to use without having to do the work 😄

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