AI2 says the new AI model beats one of the best of DeepSek

Switch to DeepSEEK. There is a new AI champion in the city – and they are Americans.

Thursday, AI2, Non-Profit AI Research Institute located in Seattle, he left a model he claimed that there were foreign claims Deepseek v3One of the Chinese AI company is one of the leading systems of DeepSeek.

AI2 model named Tulu3-405B also beats Openai GPT-4O About certain AI criteria for the internal test of AI2. Moreover, GPT-4O (and even different from DeepSeek V3), Tulu3-405B open sourceIt is all and free of all the components necessary to repeat it from scratch and Licensed as permit.

A spokesman for AI2 TechCrunch, Tulu3-405B of the laboratory, “emphasizes the potential of the United States to lead the global development of the best generative AI models in the best grade.”

“This stage is the main point for the future of the EU, the US is the main point to lead as a leader in competitive and open source models,” he said. “This launcher, AI2, Deepseek’s models, the United States offers an advanced alternative – not only in the development of the AI, but also the technical giants of the United States, which can lead to a competitive, open source AI.”

Tulu3-405B is a very large model. According to AI2, 256 GPUs consisting of 405 billion parameters consisting of 405 billion parameters require the participation of 256 GPUs. Settings are approximately a model of problem solving skills and models with more parameters generally perform better than less parameters.

According to AI2, the keys to get competitive performance with Tulu3-405B are a technique called reinforcement learning with checkable prizes. Learn to reinforce with validable rewards or RLVR, the following instructions, RLVR, RlvR models.

AI2, Benchmark Popga, 14,000 specialized knowledge from Wikipedia claims to defeat Tulu3-405B only DeepSeek V3 and GPT-4O Meta’s Llama 3.1 405b The model. Tulu3-405B, as well as in GSM8K in the highest performance of any model in the classroom, in a test with classy math word problems.

TULU3-405B Available to test AI2’s Chatbot website application and Code to train the model and make delicate in GitHub. Get in a warm case – the next benchmark comes next to the flagship AI model.