zen-multilingual

Zen Multilingual

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

Base: Qwen3-8B | Parameters: 8B | Architecture: Qwen3 | Context: 128K | License: Apache 2.0

Multilingual generation across 30+ languages: English, Chinese, Japanese, Korean, Arabic, Spanish, French, German, Portuguese, Russian, and more.

Strong at cross-lingual reasoning, code-switching, and multilingual instruction following.

Base weights: Qwen/Qwen3-8B (Apache 2.0).

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("zenlm/zen-multilingual", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-multilingual")
messages = [{"role": "user", "content": "Your domain-specific prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Credits

Built on Qwen3-8B by the Qwen Team, Alibaba Cloud, licensed under Apache 2.0. Hanzo adds identity training, agentic-data fine-tuning, and abliteration.

---

The Zen LM Family

Joint research between Hanzo AI (Techstars '17), Zoo Labs Foundation (501c3), and Lux Partners Limited.

All weights Apache 2.0. Download, run locally, fine-tune, deploy commercially.

HuggingFace · Chat · API · Docs