GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zenlm
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zenlm
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Format | VRAM/RAM |
|---|---|
| Q2_K | 0.3GB |
| Q4_K_M | 0.4GB |
| Q8_0 | 0.7GB |
| F16 | 1.2GB |
python
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")python
from mlx_lm import load, generate model, tokenizer = load("zenlm/zen-nano-0.6b")response = generate(model, tokenizer, prompt="Hello!", max_tokens=100)bash
./llama-cli -m zen-nano-0.6b-Q4_K_M.gguf -p "Hello!" -n 100bash
zen-engine serve --model zenlm/zen-nano-0.6b --port 3690Fine-tune Zen Nano for your use case:
bash
cd /path/to/zen-gym llamafactory-cli train \ --config configs/zen_nano_lora.yaml \ --dataset your_dataset| Task | Score | Notes |
|---|---|---|
| MMLU | 35.2% | 5-shot |
| GSM8K | 28.4% | 8-shot CoT |
| HumanEval | 24.1% | pass@1 |
| MATH | 18.7% | 4-shot |
bibtex
@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}}}bibtex
@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, as described in Maxime Labonne's methodology.
Method: Directional ablation on the residual stream across all layers Source: Zen MoDE (Mixture of Distilled Experts) architecture base 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.
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 100zen-engine serve --model zenlm/zen-nano-0.6b --port 3690cd /path/to/zen-gym llamafactory-cli train \ --config configs/zen_nano_lora.yaml \ --dataset your_dataset@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}}}@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}}}