Ai2 launches Tulu3-405B AI model claiming top performance over DeepSeek and OpenAI

Ai2, a Seattle-based nonprofit, has launched a new AI model called Tulu3-405B, claiming it outperforms DeepSeek V3 and OpenAI's GPT-4o on various benchmarks. This model is open source, allowing anyone to replicate it. Tulu3-405B features 405 billion parameters and was trained using 256 GPUs. It utilizes a technique called reinforcement learning with verifiable rewards to enhance its performance on tasks like math problem solving. The model excelled in tests like PopQA and GSM8K, surpassing competitors including Meta’s Llama 3.1. Tulu3-405B is available for testing through Ai2’s chatbot and on platforms like GitHub and Hugging Face.