zen-omni/model_card.md

Model Card: zen-omni

CRITICAL: Base Model Clarification

⚠️ IMPORTANT: This model is based on Qwen3-Omni, NOT Qwen2.5-32B-Instruct.

Correct Base Model Information:

• Base: Qwen3-Omni
• Repository: https://github.com/QwenLM/Qwen3-Omni
• Architecture: Multimodal Transformer (Text + Vision + Audio)
• NOT: Qwen2.5-32B or any Qwen2 variant

What is Qwen3-Omni?

Qwen3-Omni is a groundbreaking multimodal foundation model that can:

• Process text, images, and audio simultaneously
• Engage in voice conversations
• Understand visual content
• Reason across different modalities

This is fundamentally different from text-only models like Qwen2.5.

Model Architecture

zen-omni inherits Qwen3-Omni's unified architecture:

Input Modalities:
├── Text → Text Encoder
├── Image → Vision Transformer  
└── Audio → Speech Encoder
        ↓
   Multimodal Fusion
        ↓
   Unified Decoder
        ↓
   Generated Output

Key Differences from Text-Only Models

Feature	zen-omni (Qwen3-Omni based)	Qwen2.5 Models
Modalities	Text + Vision + Audio	Text only
Architecture	Multimodal Transformer	Text Transformer
Vision Understanding	✅ Native	❌ Not supported
Audio Processing	✅ Native	❌ Not supported
Use Cases	Multimodal AI	Text generation

Capabilities

Based on Qwen3-Omni's features:

• Voice Interaction: Real-time speech conversations
• Visual Understanding: Analyze images and videos
• Audio Processing: Transcribe and understand audio
• Cross-Modal Reasoning: Connect information across modalities

Technical Specifications

base_model: Qwen3-Omni
modalities:
  - text
  - image  
  - audio
model_type: multimodal-transformer
parameters: ~7B
context_length: 32768

Important Notes

1. This is NOT a scaled-up Qwen2.5 model
2. This is NOT just a text model
3. This IS a multimodal model based on Qwen3-Omni
4. Requires multimodal-compatible inference code

Citation

@article{zen-omni,
  title={Qwen3-Omni: Unified Multimodal Intelligence},
  author={Qwen Team},
  year={2024}
}

---

Last Updated: November 2024
Correction: Model base changed from incorrect Qwen2.5 to correct Qwen3-Omni