Key Features
• Advanced Reasoning
-1. Optimized for math, coding, and multi-step problem-solving
-2. Uses reinforcement learning (RL) for step-by-step logical thinking
-3. Competitive with DeepSeek-R1 and o1-mini on benchmarks
• Efficient Design
-1. 32 billion parameters, dense architecture for high performance
-2. Runs locally on consumer GPUs (e.g., 24GB VRAM)
• Extended Context
-1. 131,072 token context window for long inputs
-2. Handles up to 300 pages of text or complex codebases
• Performance Highlights
-1. Scores 79.5% on AIME24 (math), 63.4% on LiveCodeBench (coding)
-2. Outperforms o1-mini in function calling (66.4% on BFCL)
• Developer-Friendly
-1. Open-source under Apache 2.0 license on Hugging Face
-2. Supports tools, structured outputs, and APIs like Groq
• Accessibility
-1. Free via Qwen Chat or local deployment
-2. Quantized versions (e.g., 4-bit AWQ) for lower resource use
