Key Features
β’ Advanced Reasoning
-1. Outperforms GPT-4o on GPQA (56.1%) and MATH (80.4%) benchmarks
-2. Excels in STEM tasks, including math competition problems (AMC 10/12: 91.8%)
-3. Uses synthetic data and rejection sampling for enhanced logic
β’ Compact Efficiency
-1. 14 billion parameters, 5x smaller than Llama 3.3 70B
-2. Runs on edge devices with 4-bit quantization (e.g., 11GB VRAM)
-3. Low-latency inference for resource-constrained settings
β’ Extended Context
-1. 16,000 token context window, up from 4K mid-training
-2. Handles long-form text and complex workflows effectively
β’ Performance Highlights
-1. Matches or beats GPT-4o-mini and Llama 3.3 70B in reasoning tasks
-2. Strong coding skills with 82.6% on HumanEval
β’ Open-Source Accessibility
-1. Available under MIT License on Hugging Face and Azure AI Foundry
-2. Supports fine-tuning with LoRA and tools like vLLM
β’ Practical Applications
-1. Ideal for research, education, and enterprise AI solutions
-2. Powers local deployment and cost-effective workflows
