Tülu 3 - Open AI Model That Beats DeepSeek v3 and Challenges GPT-4o

Ai2's new open AI model Tülu 3 405B challenges GPT-4o and outperforms DeepSeek v3 with its innovative post-training technique. With full transparency into code and data, Ai2 sets a new standard for open AI. Is this the start of a new era where open models seriously compete with the big AI giants?

Share
Tülu 3 - Open AI Model That Beats DeepSeek v3 and Challenges GPT-4o

The open AI market takes a new step forward as the Allen Institute for AI (Ai2) launches Tülu 3 405B, an open language model with 405 billion parameters. The model performs on par with OpenAI's GPT-4o and outperforms DeepSeek v3 in several key tests.

Ai2 has previously claimed to challenge the market's leading models, and back in November 2024, the company launched the first versions of Tülu 3 in 8- and 70-billion sizes. The latest update, however, takes the model to a whole new level.

Groundbreaking Post-Training Improves AI Performance

Tülu 3 405B's main innovation lies in its advanced post-training technique, where the model undergoes a unique combination of:

Supervised fine-tuning (supervised fine-tuning)

Preference learning (preference learning)

Reinforcement learning from verifiable rewards (RLVR)

The latter method is especially groundbreaking, letting the model be adjusted based on verifiable outcomes, such as math problems with an objectively correct answer. This gives an AI with both higher accuracy and better safety characteristics.

RLVR - Revolutionizing AI Training

Ai2's RLVR system stands out by using verifiable rewards to guide the model's learning, unlike traditional methods that rely solely on human preferences. The combination of RLVR and optimized computing infrastructure has enabled:

Efficient parallel processing across 256 GPUs

Weight synchronization for optimized performance

Balanced workload distribution across 32 nodes

Integrated vLLM deployment with 16-way tensor parallelism

The RLVR system has proven especially effective in larger models, suggesting the method could deliver even better results in future, even more extensive AI models.

Performance Comparison: Tülu 3 405B vs GPT-4o and DeepSeek v3

Tülu 3 405B has been tested against ten leading AI benchmark tests, including safety evaluations. The model performed better than DeepSeek v3, but slightly below GPT-4o:

Tülu 3 405B RLVR: 80.7 points
DeepSeek v3: 75.9 points
GPT-4o: 81.6 points

Although GPT-4o still holds a slight lead, the results show that Tülu 3 405B is one of the most competitive open models on the market today.

Open Source for Real - Ai2's Unique Strategy

One of the biggest differences between Tülu 3 and other open models like DeepSeek and Meta's Llama 3.1 is the degree of transparency.

While DeepSeek and Llama 3.1 only share certain parts of their code and weight data, Ai2 offers a fully open solution:

Complete training data

All code and infrastructure

The ability to do custom training and fine-tuning

"We don't use any closed datasets," says Hannaneh Hajishirzi, senior director of NLP Research at Ai2.

For those who want to try the model, it's available via Ai2's Tülu 3 platform, where both training data and code are freely accessible.

A New Era for Open AI

The launch of Tülu 3 405B marks an important step in the development of open AI models. By combining transparency, innovative training methods, and high performance, Ai2 has established one of the strongest challengers to the big, closed AI models from OpenAI and DeepSeek.

The question now is - how will the tech giants respond to the open AI revolution?