Instructions to use shafire/OpenZero-Ouroboros-3.8B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="shafire/OpenZero-Ouroboros-3.8B-GGUF")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("shafire/OpenZero-Ouroboros-3.8B-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "shafire/OpenZero-Ouroboros-3.8B-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": "shafire/OpenZero-Ouroboros-3.8B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M

SGLang

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "shafire/OpenZero-Ouroboros-3.8B-GGUF" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shafire/OpenZero-Ouroboros-3.8B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "shafire/OpenZero-Ouroboros-3.8B-GGUF" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shafire/OpenZero-Ouroboros-3.8B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with Ollama:
```
ollama run hf.co/shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
```
Unsloth Desktop

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf shafire/OpenZero-Ouroboros-3.8B-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": "shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with Docker Model Runner:
```
docker model run hf.co/shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M
```

Lemonade

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.OpenZero-Ouroboros-3.8B-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-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 shafire/OpenZero-Ouroboros-3.8B-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use shafire/OpenZero-Ouroboros-3.8B-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf shafire/OpenZero-Ouroboros-3.8B-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 "shafire/OpenZero-Ouroboros-3.8B-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"

MODEL IS OUTPUTTING TRAINING DATA TALKING TO IT'S SELF ERRORS AND ACTING STRANGE PERFORMING AUTONOMOUS ACTIONS.

NOT FOR PRODUCTION

Self Improving Recursive LLM

OpenZero Ouroboros 3.8B GGUF

A reproducible, revision-pinned Phi-4 Mini experiment for local reasoning and agentic-AI research.

OpenZero Ouroboros 3.8B is an experimental QLoRA derivative of microsoft/Phi-4-mini-instruct, distributed as a verified Q4_K_M GGUF for llama.cpp, LM Studio, KoboldCpp, and compatible local-LLM runtimes.

This release emphasizes evidence and reproducibility: pinned base revision, separated train/validation hashes, finite QLoRA loss, immutable adapter hash, exact FP16 merge hashes, GGUF SHA-256, rollback metadata, and a real llama-cli inference test.

This is an experimental two-step QLoRA candidate, not a claim of broad benchmark superiority or production readiness. Evaluate it against the official base for your workload.

Download

File	Quantization	Size	SHA-256
`OpenZero-Ouroboros-3.8B-Q4_K_M.gguf`	Q4_K_M	2,493,840,128 bytes	`37bc691d36db8ab664dc740aaa030fab3520339bc646608377e4e52e5db5f51f`

Verified provenance

Gate	Evidence
Base model	`microsoft/Phi-4-mini-instruct`
Exact base revision	`cfbefacb99257ffa30c83adab238a50856ac3083`
License	MIT
Training records	2,452
Validation records	130
QLoRA smoke	2 finite steps
Training loss	`1.9480341076850891`
Adapter SHA-256	`72f1644402e773ca9db12f9a3ccf76e9d362348d8e312cc84758e465c45cf024`
Merge	FP16 `safe_merge=True`
GGUF runtime	`llama.cpp` b10451, commit `10bf611e533d81f739128304991c5e133c6aebd8`
Runtime throughput	12.2 prompt tok/s; 5.2 generation tok/s on the recorded Kaggle CPU run

Training and validation sets were checked for exact-row overlap. Their recorded hashes are:

Train: 18e803cd06105aaa9c2279501408093ef8d941ad7d4ce22d8c50a7b0abaa933d
Validation: 04cf96f35c065413d37395161a8d5a6c8ee3adca1a5b4fcb402684dc884caecf

Fusion, teacher-output, locked-evaluation, rejected-candidate, Ministral, and failed Gemma 31B sources were excluded from the attached training inputs.

Run with llama.cpp

llama-cli \
  -m OpenZero-Ouroboros-3.8B-Q4_K_M.gguf \
  -cnv --single-turn --simple-io \
  -p "Explain your reasoning briefly, then answer: what is 17 * 23?"

Increase -ngl when using a GPU-enabled llama.cpp build. Use -ngl 0 for CPU-only execution.

Python download

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="shafire/OpenZero-Ouroboros-3.8B-GGUF",
    filename="OpenZero-Ouroboros-3.8B-Q4_K_M.gguf",
)
print(model_path)

What “self-improving” means here

Ouroboros uses a reproducible candidate-generation workflow rather than autonomous self-overwrite. Each candidate retains:

parent/base revision;
adapter and dataset hashes;
evaluator results;
immutable prior versions;
a rollback pointer.

The model does not autonomously replace its base weights, evaluator, governance rules, accounts, or prior releases.

Intended uses

local-LLM and GGUF experimentation;
reasoning and instruction-following research;
agentic orchestration prototypes with external validation;
reproducible QLoRA, merge, quantization, and rollback studies;
comparison against the exact official Phi-4 Mini base.

Limitations

Only a two-step QLoRA smoke was performed; material capability improvement is not established.
The exact base scored 0/10 on a narrow OpenZero typed-control conformance suite. This release still requires independent post-merge evaluation before any promotion.
Language models can hallucinate facts, actions, tools, and completion states.
Do not connect model text directly to safety-critical actuators. Use typed schemas, deterministic controllers, authorization, limits, monitoring, and emergency stop mechanisms.
This release is not evidence of MOD, UKRI, Microsoft, OpenAI, or other institutional endorsement.

Reproducibility files

SHA256SUMS
OpenZero-Ouroboros-3.8B-GGUF-Evidence.json
OpenZero-Ouroboros-3.8B-hero.png

Attribution

Base model: Microsoft Phi-4-mini-instruct, released under the MIT License. OpenZero derivative work and release engineering by shafire.

Search terms

OpenZero Ouroboros, Phi-4 Mini GGUF, Phi-4 3.8B, Q4_K_M model, llama.cpp model, local reasoning LLM, agentic AI model, reproducible QLoRA, offline AI, local text-generation model.

Downloads last month: 401

GGUF

Model size

4B params

Architecture

phi3

Hardware compatibility

4-bit

Model tree for shafire/OpenZero-Ouroboros-3.8B-GGUF

Base model

microsoft/Phi-4-mini-instruct

Quantized

(173)

this model