Available Formats

This model is available in multiple formats for different platforms:

SafeTensors (Base Format)

• Standard HuggingFace format
• Compatible with Transformers library
• Use for training and fine-tuning

MLX Format (Apple Silicon Optimized)

• /mlx/ - Full precision MLX format
• /mlx-4bit/ - 4-bit quantized (fastest on Mac)

GGUF Format (Coming Soon)

• Will be added for llama.cpp compatibility
• CPU-optimized for all platforms

Quick Start

Using Transformers

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-guard-stream-4b")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-guard-stream-4b")

Using MLX (Apple Silicon)

python
Copy code
from mlx_lm import load, generate

# Load 4-bit model (fastest)
model, tokenizer = load("zenlm/zen-guard-stream-4b", adapter_path="mlx-4bit")

# Generate
response = generate(model, tokenizer, prompt="Your prompt", max_tokens=256)
print(response)

Using llama.cpp (GGUF - Coming Soon)

bash
Copy code
llama-cli -m gguf/zen-guard-stream-q4_k_m.gguf -p "Your prompt"

Training with Zoo-Gym

bash
Copy code
pip install zoo-gym
zoo-gym train --model zenlm/zen-guard-stream-4b --data your_data.jsonl

Model Details

• Architecture: Zen Guard architecture
• Training: Zoo-Gym with RAIS (Recursive AI Self-Improvement System)
• License: Apache 2.0
• Partnership: Hanzo AI x Zoo Labs Foundation

Citation

bibtex
Copy code
@misc{zen_zen_guard_stream_2025,
    title={Zen Guard Stream v1.0.1},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    version={1.0.1}
}