Summarize this content to 2000 words in 6 paragraphs
The Allen Institute for AI (Ai2) released a supersized version of its Tülu 3 AI model, aiming to further advance the field of open-source artificial intelligence and demonstrate its own techniques for enhancing the capabilities of AI models.
The Seattle-based institute said the new Tülu 3 405B model, announced Thursday morning, rivals or exceeds the performance of OpenAI’s GPT 4o and DeepSeek v3 — the open-source model from China that has grabbed the attention of the tech world and challenged the premise of the industry’s massive AI investments.
Ai2 said its new model shows that the U.S. can produce competitive, open-source artificial intelligence, free from the influence of big tech companies.
The new Tülu version has 405 billion parameters, compared with 70 billion parameters in the largest prior version, significantly increasing the model’s ability to capture the complexity of patterns and relationships in its training data.
A chatbot demonstration of the Tülu 3 405B model is now available, along with an updated research paper, and access to the underlying code on GitHub.
Like DeepSeek, the Ai2 Tülu project focuses on post-training, the process of refining a language model to enhance its capabilities and make it suitable for specific tasks.
Tülu 3 405B is “the largest fully open-source post-trained model to date,” Ai2 said.
Ai2 says Tülu 3 rivals or surpasses other AI models on key benchmarks. (Ai2 chart, click to enlarge.)
Ai2’s post-training approach includes a technique called Reinforcement Learning from Verifiable Rewards, or RLVR, which involves training a model by rewarding or penalizing it based on whether its responses are correct for objectively verifiable tasks like solving math problems and following instructions.
Ai2 first demonstrated the effectiveness of its RLVR approach with the release of its prior Tülu 3 models in November. The latest release demonstrates the ability to successfully deploy RLVR at a much larger scale, Ai2 said.
“The primary objective of this release was to stress-test our novel RLVR approach and training infrastructure at large scales and extend the Tülu 3 recipe to the Llama-405B base model,” the Ai2 Tülu team explained in a blog post.
Even at that scale, the “training pipeline proved robust,” they wrote.
DeepSeek likewise used reinforcement learning as a core component of post-training, but with some different techniques, including zero or minimal supervised fine-tuning, a preliminary step that uses labeled data. DeepSeek also showed the value of distilling existing models into smaller versions to improve performance.
The resulting efficiencies raised questions about the investments being made across the industry to build out new infrastructure for training AI models, sending major tech stocks plummeting earlier this week.
Ai2 was founded in 2014 by the late Microsoft co-founder Paul Allen. It has ties to the University of Washington’s Allen School of Computer Science & Engineering, including several UW faculty who are also Ai2 research leaders.
Last year, Ai2 released new multimodal artificial intelligence model, dubbed Molmo, that works with visual data in novel ways. Ai2 released its Open Language Model, or OLMo, in 2023, as part of a larger effort to bring more transparency to AI models.
