Instructions to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="illegalcall/Qwen3.5-9B-RYS-0-4-GGUF",
	filename="Qwen3.5-9B-RYS-0-4.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
# Run inference directly in the terminal:
llama-cli -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
# Run inference directly in the terminal:
llama-cli -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

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 illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
# Run inference directly in the terminal:
./llama-cli -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

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 illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Use Docker

docker model run hf.co/illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

LM Studio
Jan

vLLM

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "illegalcall/Qwen3.5-9B-RYS-0-4-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": "illegalcall/Qwen3.5-9B-RYS-0-4-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Ollama
How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Ollama:
```
ollama run hf.co/illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
```

Unsloth Studio new

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for illegalcall/Qwen3.5-9B-RYS-0-4-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for illegalcall/Qwen3.5-9B-RYS-0-4-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for illegalcall/Qwen3.5-9B-RYS-0-4-GGUF to start chatting

Pi new

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "illegalcall/Qwen3.5-9B-RYS-0-4-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

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 illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Run Hermes

hermes

Docker Model Runner
How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Docker Model Runner:
```
docker model run hf.co/illegalcall/Qwen3.5-9B-RYS-0-4-GGUF
```

Lemonade

How to use illegalcall/Qwen3.5-9B-RYS-0-4-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Run and chat with the model

lemonade run user.Qwen3.5-9B-RYS-0-4-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Qwen3.5-9B-RYS-0-4 (GGUF)

An experimental model created by duplicating layers 0-3 in Qwen3.5-9B using the RYS (Repeat Your Self) method.

What This Is

The base Qwen3.5-9B model (32 layers) with layers 0-3 physically duplicated, creating a 36-layer model. The forward pass executes: layers 0,1,2,3 → 0,1,2,3 (again) → 4,5,...,31.

No training, no fine-tuning, no weight changes — just architectural routing.

Method

Built using llm-circuit-finder:

python layer_path.py Qwen3.5-9B-Q4_K_M.gguf \
    Qwen3.5-9B-RYS-0-4.gguf \
    -p "0..3,0,1,2,3,4..31" -v

Architecture Discovery

Qwen 3.5 uses a hybrid [DeltaNet, DeltaNet, DeltaNet, Attention] repeating pattern (4-layer cycles). Layer duplication must use block sizes that are multiples of 4 to preserve this pattern — block size 3 crashes with "missing tensor" errors.

This is the first known application of the RYS method to a hybrid DeltaNet/Attention architecture.

Fair Evaluation Results (max_tokens=4096, no think-tag stripping)

Model	Code Gen	Hallucination Resistance	Reasoning	Overall
Baseline (32 layers)	80%	40%	100%	73.3%
RYS (0,4) (36 layers)	60%	80%	100%	80.0%

Tested with 15 questions (5 code generation, 5 hallucination detection, 5 reasoning). Both models evaluated under identical conditions.

Key finding: The improvement is primarily in hallucination resistance (+40 percentage points). Code generation shows a tradeoff (-20 percentage points). Reasoning is unchanged.

Important caveats:

Sample size is small (5 questions per category) — results need validation with larger benchmarks
The improvement may be related to response generation behavior rather than capability differences
The original RYS method was validated on standard transformer architectures (72B+ models); Qwen 3.5's hybrid DeltaNet architecture is untested territory
Side-by-side testing on harder hallucination prompts showed identical responses from both models

Circuit Map of Qwen 3.5-9B

Full sweep results (7 configurations tested):

Config	Code	Hallucination	Reasoning	Overall	vs Baseline
Baseline	80%	40%	100%	73.3%	—
(0,4)	60%	80%	100%	80.0%	+6.67%
(4,8)	80%	60%	80%	73.3%	+0.00%
(8,12)	0%	40%	80%	40.0%	-33.33%
(12,16)	0%	60%	80%	46.7%	-26.67%
(16,20)	0%	40%	100%	46.7%	-26.67%
(20,24)	60%	60%	100%	73.3%	+0.00%

Usage

# With llama.cpp
llama-server -m Qwen3.5-9B-RYS-0-4.gguf -c 8192 -ngl 99

# With Ollama (create a Modelfile)
echo 'FROM ./Qwen3.5-9B-RYS-0-4.gguf' > Modelfile
ollama create qwen35-rys -f Modelfile
ollama run qwen35-rys

Specifications

Base model: Qwen/Qwen3.5-9B
Quantization: Q4_K_M
Original layers: 32
Modified layers: 36 (layers 0-3 duplicated)
File size: ~6.2 GB
Extra VRAM: ~0.5 GB over base model
Inference overhead: ~12% slower

References

License

Apache 2.0 (same as base model)

Downloads last month: 28

GGUF

Model size

10B params

Architecture

qwen35

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Quantized

(222)

this model

Paper for illegalcall/Qwen3.5-9B-RYS-0-4-GGUF

H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs

Paper • 2512.01797 • Published Dec 1, 2025 • 9