American AI research institute Ai2 has just dropped a bombshell. The Seattle nonprofit released an AI model reportedly overtaking one of the best from Chinese AI giant, DeepSeek.
According to internal tests run at Ai2, the new model, Tulu3-405B, surpasses DeepSeek’s V3 as well as OpenAI’s GPT-4o on specific AI benchmarks. The standout feature? Unlike its competitors, Tulu3-405B is open-source, with all its components freely available, fostering openness and collaboration.
An Ai2 spokesperson emphasised the potential for this model to elevate the United States’ AI game. Tulu3-405B validates the country’s capacity to spur the global evolution of high-calibre generative AI models.
Tulu3-405B isn’t just another model. It’s a mammoth, grasping 405 billion parameters. Ai2 revealed that training this giant took a spread of 256 GPUs working simultaneously. In the AI world, increasing parameters can significantly boost model performance.
Ai2 accredits much of Tulu3-405B’s stellar performance to the application of a technique called reinforcement learning with verifiable rewards. This technique teaches tasks providing ‘verifiable’ outcomes, such as mathematical problem solving.
On the benchmark board, Tulu3-405B outwitted not just DeepSeek V3 and GPT-4o, but also Meta’s model Llama 3.1 405B on PopQA, a set of over 14,000 knowledge-based questions. Moreover, this AI model proved its mettle in resolving grade school-level math word problems.
Curious to test this groundbreaking model? Ai2’s chatbot web app has got you covered. You can also train Tulu3-405B using its code available on GitHub and the Hugging Face AI development platform. So, gear up and tap into the potential of this AI marvel!
Original source: Read the full article on TechCrunch
