zenlm/zen-omni-30b-thinking

Zen Omni 30B Thinking

Advanced multimodal model with thinking and audio capabilities from the Zen family.

Model Details

• Architecture: Qwen2-based with multimodal extensions
• Parameters: 31.7B
• Context Length: 32,768 tokens
• Modalities: Text, Audio, Thinking
• Hidden Size: 5,120
• Layers: 64
• Attention Heads: 40
• Developer: Hanzo AI

Features

• Thinking Module: Chain-of-thought reasoning capabilities
• Audio Tower: Audio processing and understanding
• Multimodal Integration: Seamless text and audio processing

Usage

PyTorch

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-omni-30b-thinking")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-omni-30b-thinking")

# Generate text
prompt = "Explain quantum computing"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Available Formats

• PyTorch: Default safetensors format (16 shards)
• GGUF: Coming soon
• MLX: Coming soon

Hardware Requirements

• VRAM: ~64GB for full precision
• RAM: 128GB recommended
• Storage: ~60GB for model files

Training

Fine-tuned with:

• Zen identity
• Multimodal understanding
• Chain-of-thought reasoning
• Audio processing capabilities

Model Components

• thinker.*: Thinking/reasoning module
• audio_tower.*: Audio processing layers
• Standard transformer layers for text generation

License

Apache 2.0