zen-family

Superseded — archived checkpoint. This is the v1.0.1 family overview; its weights are a Qwen3-8B fine-tune. See the current lineup at huggingface.co/zenlm. Weights are kept available for reproducibility.

Zen AI Model Family (v1.0.1, archived)

Base model & attribution

Fine-tuned from Qwen/Qwen3-8B (Apache-2.0) with Hanzo identity + agentic-data training + abliteration.

Zen fine-tunes the best open-weight model of each era. Across this family that is the Qwen3 line from Alibaba Cloud — base, VL, Omni, Embedding, Reranker, TTS, ASR, Guard, Coder — with media models on permissive bases (Wan, FLUX, TRELLIS, YuE). Hanzo adds only identity training, agentic-data fine-tuning, and abliteration. There is no from-scratch Zen model.

• Upstream model: Qwen/Qwen3-8B
• Upstream project: Qwen (Alibaba Cloud)
• Upstream license: Apache-2.0

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-family")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-family")

messages = [{"role": "user", "content": "Hi, what can you help me with?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Models

Browse the current, canonical lineup at huggingface.co/zenlm.

Citation

@misc{zen_family_2025,
    title  = {Zen AI Model Family},
    author = {Hanzo AI and Zoo Labs Foundation},
    year   = {2025},
    url    = {https://huggingface.co/zenlm}
}

License

Apache-2.0, inherited from the upstream Qwen3-8B base. See LICENSE and NOTICE.

---

Built by Hanzo AI and Zoo Labs Foundation.