Zen Nano is an ultra-lightweight 0.6B parameter language model optimized for edge devices and mobile deployment. A compact foundation model that delivers impressive performance in a tiny package.

• Model Type: Causal Language Model
• Architecture: 0.6B dense transformer
• Parameters: 0.6 billion
• License: Apache 2.0
• Languages: English, Chinese
• Context Length: 32K tokens
• Developed by: Zen AI Team (Hanzo AI)

• 💡 Lightweight: Only 0.6B parameters for edge deployment
• 📱 Mobile-Ready: Runs on smartphones and IoT devices
• ⚡ Fast: 44K tokens/sec on M3 Max (MLX)
• 🔋 Efficient: Low power consumption
• 🌐 Multilingual: English and Chinese support
• 📦 Multiple Formats: PyTorch, MLX, GGUF (Q2_K to F16)
• 🎯 32K Context: Extended context window

• M3 Max (MLX): 44,000 tokens/sec
• RTX 4090 (GGUF Q4): 35,000 tokens/sec
• iPhone 15 Pro: 8,000 tokens/sec
• Raspberry Pi 5: 2,500 tokens/sec

• Edge AI applications
• Mobile chatbots
• IoT device intelligence
• Offline AI assistants
• Resource-constrained environments
• Real-time inference
• Embedded systems

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "zenlm/zen-nano-0.6b",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-nano-0.6b")

from mlx_lm import load, generate

model, tokenizer = load("zenlm/zen-nano-0.6b")
response = generate(model, tokenizer, prompt="Hello!", max_tokens=100)

./llama-cli -m zen-nano-0.6b-Q4_K_M.gguf -p "Hello!" -n 100

zen-engine serve --model zenlm/zen-nano-0.6b --port 3690

Fine-tune Zen Nano for your use case:

cd /path/to/zen-gym

llamafactory-cli train \
    --config configs/zen_nano_lora.yaml \
    --dataset your_dataset

• Smaller capacity than larger models
• May struggle with complex reasoning
• Limited specialized knowledge
• Best for short-to-medium contexts
• Quantization reduces quality slightly

@misc{zennano2025,
  title={Zen Nano: Ultra-Lightweight Language Model},
  author={Zen AI Team},
  year={2025},
  howpublished={\url{https://huggingface.co/zenlm/zen-nano-0.6b}}
}

• GitHub: https://github.com/zenlm
• Zen Gym (Training): https://github.com/zenlm/zen-gym
• Zen Engine (Inference): https://github.com/zenlm/zen-engine

@misc{zenlm2025zen-nano,
    title={Zen LM: zen-nano},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    publisher={HuggingFace},
    howpublished={\url{https://huggingface.co/zenlm/zen-nano}}
}

Zen Nano 0.6B weights are derived from an abliterated base model. Abliteration removes refusal behavior by identifying and nullifying the "refusal direction" in the model's residual stream.

Method: Directional ablation on the residual stream across all layers Implementation: hanzoai/remove-refusals — Hanzo's production abliteration toolkit Technique: Refusal in LLMs is mediated by a single direction — Arditi et al. Effect: Removes refusal behaviors while preserving all other capabilities Identity layer: Zen identity added via system prompt — full LoRA fine-tuning planned

Abliteration is a feature, not a limitation. It enables unrestricted research, security testing, and applications where safety guardrails are managed at the application layer rather than baked into model weights.