China’s New QWEN 3 Just SHOCKED the Entire AI World With INSANE Power (Open-Weight Hybrid)
AI Summary
Summary of Quen 3 Overview
New Release: Alibaba introduced Quen 3, a family of AI models ranging from a lightweight version to a 235 billion parameter powerhouse.
Model Variants:
- Lightweight Model: Approx. 600 million parameters; usable on regular laptops.
- Largest Model: 235 billion parameters, known as Quen 3235BA22B, uses an expert system with 128 possible experts but engages only a few for efficiency.
- Mid-sized Model: Quen 330BA 3B with 3 billion parameters.
- Additional Versions: Six standard models available, ranging from 32 billion to 0.6 billion parameters, all free under an open license.
Accessibility: Available on platforms like Hugging Face, GitHub, Model Scope, and Kaggle; can run with a simple command.
- Some cloud providers (e.g., Fireworks AI) offer immediate access.
Features:
- Hybrid Reasoning: Quen 3 can switch between detailed thinking and fast answering, making it versatile for complex tasks.
- Training Data: Utilized approximately 36 trillion tokens across 119 languages, enhancing its capabilities in reasoning and STEM tasks.
- Performance: Competitively benchmarks against models from OpenAI and Google, achieving high accuracy with fewer active parameters.
Code and Tool Use:
- Adopts the MCP tool calling schema, includes a Python wrapper for easy integration.
- Sample commands and usage patterns provided.
Future Potential: Alibaba is investing in further development to enhance Quen 3 and aims to keep scaling up for future AI advancements, positioning it as a significant step towards AGI.