Tulu 3 405b: A New Open-Source AI Model Challenging DeepSeek and OpenAI
The field of Artificial Intelligence (AI) is rapidly evolving, with new models and advancements emerging constantly. One of the latest contenders in the large language model (LLM) space is Tulu 3 405b, an open-source model developed by the Allen Institute for AI (AI2). This massive model boasts 405 billion parameters and has shown impressive performance on various benchmarks, challenging established players like DeepSeek and OpenAI. This article provides a general overview of Tulu 3 405b, its capabilities, and how it compares to DeepSeek and OpenAI.
Tulu 3 405b: An Overview
Tulu 3 405b is an open-source LLM developed by AI2, built upon Meta’s Llama 3.1 architecture 1. It is the first model of its kind where all the internal workings are publicly accessible, to successfully apply a fully open post-training recipe at a 405-billion parameter scale 3. The model is designed for improved performance across diverse tasks, including knowledge, reasoning, mathematics, coding, instruction following, and safety 4.
One of the key innovations of Tulu 3 405b is its novel reinforcement learning approach known as Reinforcement Learning with Verifiable Rewards (RLVR) 3. This method significantly improves model performance in specialized tasks by ensuring that rewards are based on verifiable outcomes rather than subjective feedback 3. For example, when the model attempts a math problem, it receives a reward only if it produces the correct solution 2. This approach helps Tulu 3 405b learn to be precise and efficient, prioritizing accuracy over plausible-sounding but incorrect responses 2.
Training a model of this scale posed several challenges for the AI2 engineers. Training Tulu 3 405b required 32 compute nodes with 256 GPUs working together, and each training step took 35 minutes 5. To manage the computational demands, the team had to employ workarounds, such as using a smaller helper model 5.
A key aspect of Tulu 3 405b is its open-source nature. Unlike many powerful models that are locked behind corporate paywalls, Tulu 3 405b allows for greater transparency and accessibility in AI research 6. This open approach has the potential to accelerate innovation across the field by enabling researchers to build on verified methods and allowing developers to focus on improvements rather than starting from scratch 2.
DeepSeek: An Open-Source Challenger
DeepSeek is a Chinese AI company that has gained significant attention for its open-source LLMs 7. The company focuses on developing models that rival or surpass existing industry leaders in both performance and cost-efficiency 8. DeepSeek’s models are designed to be modular and transparent, with a strong emphasis on explainability and adaptability 9.
DeepSeek has released a series of impressive models, including DeepSeek-V3 and DeepSeek-R1 10. DeepSeek-V3 boasts 671 billion parameters and was trained on a dataset of 14.8 trillion tokens 8. It outperforms models like Llama 3.1 and Qwen 2.5, while matching the capabilities of GPT-4o and Claude 3.5 Sonnet 8. DeepSeek-R1 focuses on logical inference, mathematical reasoning, and real-time problem-solving 8. It leverages a hybrid approach combining reinforcement learning with supervised fine-tuning 9, employing group relative policy optimization (GRPO) to enhance reasoning capabilities 8. DeepSeek-R1 also utilizes a “chain-of-thought” approach, which allows the model to explicitly reason out the prompt before answering, similar to how humans might approach problem-solving.
DeepSeek’s success is particularly noteworthy given the US sanctions on China’s access to advanced chips 11. This achievement challenges the assumption of US dominance in AI and highlights the potential for innovation in alternative environments 11.
One of the key innovations of DeepSeek is its “mixture of experts” architecture. This architecture allows the model to activate only a subset of its parameters for different tasks, significantly improving efficiency and reducing computational costs. This approach contributes to DeepSeek’s impressive chat-time efficiency, allowing it to provide quick and accurate responses while minimizing resource usage.
DeepSeek’s approach to AI development also includes a unique recruitment strategy. The company actively recruits young AI researchers from top Chinese universities and hires individuals from outside the computer science field to diversify its models’ knowledge and abilities.
Furthermore, DeepSeek’s use of synthetic training data, particularly its use of o1-generated “thinking” scripts to train DeepSeek-R1, highlights the potential of this approach in advancing AI. This method could lead to more efficient and effective training processes in the future, potentially reducing the reliance on massive datasets and expensive hardware.
OpenAI: The Industry Leader
OpenAI is a leading AI research and deployment company that has been at the forefront of LLM development. Its GPT models, particularly GPT-4o and o1, are known for their advanced natural language processing (NLP) capabilities, enabling them to understand and generate human-like text with remarkable accuracy 9. OpenAI’s models excel in various tasks, including creative content generation, code generation, and general problem-solving 9. OpenAI’s models are based on the transformer architecture 12.
OpenAI has invested heavily in AI research and development, with its models trained on massive datasets and utilizing cutting-edge hardware. However, OpenAI’s models are not open source, and access to their most powerful versions can be expensive 11.
Comparing Tulu 3 405b, DeepSeek, and OpenAI
| Feature | Tulu 3 405b | DeepSeek | OpenAI |
|---|---|---|---|
| Architecture | Llama 3.1 based 2 | Mixture of Experts (MoE) 13 | Transformer based 12 |
| Training Data | Publicly available datasets, synthetic data, and human-created content 14 | 14.8 trillion tokens for V3 8 | Massive datasets, including web data and code 4 |
| Key Differentiator | RLVR training method 1 | Open-source, modular, and adaptable models 9 | Advanced NLP capabilities and strong performance 9 |
| Strengths | Strong performance in mathematical reasoning and safety 1, open-source 1 | Cost-effective performance 8, open-source 7 | High accuracy and versatility 9 |
| Weaknesses | Limited safety training 4 | Potential bias due to Chinese censorship laws | Closed-source and expensive 11 |
Tulu 3 405b stands out for its RLVR training method, which allows it to excel in tasks with verifiable outcomes, such as math and coding 1. It also demonstrates a consistent edge over DeepSeek V3, especially in safety benchmarks 1. DeepSeek differentiates itself by offering open-source models with competitive performance at a lower cost 8. OpenAI, while not open source, remains a leader in the field due to its strong performance and advanced NLP capabilities 9.
Getting Started with Tulu 3 405b
For those interested in experimenting with Tulu 3 405b, there are several ways to get started. Users can easily load the model using Hugging Face’s Transformers library 14. Here’s a code snippet demonstrating how to load the model for text generation tasks:
Python
from transformers import AutoModelForCausalLM
tulu_model = AutoModelForCausalLM.from_pretrained(“allenai/Llama-3.1-Tulu-3-405B”)
Alternatively, users can interact directly with the model through the AI2 Playground, a free web demo that allows users to enter text prompts and test the model’s capabilities 14.
Conclusion
Tulu 3 405b is a significant development in the open-source AI landscape. Its competitive performance, innovative training method, and open-source nature make it a strong contender against established players like DeepSeek and OpenAI. The emergence of Tulu 3 405b has several implications for the ongoing AI race between the US and China. It demonstrates that high-performing AI models can be developed outside of the traditional tech giants and that open-source initiatives can play a crucial role in driving innovation and accessibility in AI.
The increasing availability of open-source models like Tulu 3 405b and DeepSeek has the potential to democratize AI technology, making it more accessible to researchers, developers, and businesses worldwide. This could lead to a more diverse and competitive AI landscape, with new applications and advancements emerging from unexpected sources.
As the AI field continues to evolve, it will be interesting to see how Tulu 3 405b and other open-source models shape the future of AI development, accessibility, and the future of the AI industry.
Works cited
- Ai2 Launches Tülu3-405B Model, Scales Reinforcement Learning for Open Source AI, accessed February 1, 2025, https://www.aiwire.net/2025/01/31/ai2-launches-tulu3-405b-model-scales-reinforcement-learning-for-open-source-ai/
- Allen AI’s Tülu 3 Just Became DeepSeek’s Unexpected Rival – Unite.AI, accessed February 1, 2025, https://www.unite.ai/allen-ais-tulu-3-just-became-deepseeks-unexpected-rival/
- The Allen Institute for AI (AI2) Releases Tülu 3 405B: Scaling Open-Weight Post-Training with Reinforcement Learning from Verifiable Rewards (RLVR) to Surpass DeepSeek V3 and GPT-4o in Key Benchmarks – Reddit, accessed February 1, 2025, https://www.reddit.com/r/machinelearningnews/comments/1ieoxoa/the_allen_institute_for_ai_ai2_releases_t%C3%BClu_3/
- Tulu | Ai2 – Allen Institute for Artificial Intelligence, accessed February 1, 2025, https://allenai.org/tulu
- Allen AI claims its new Tülu 3 405B open source model rivals top performers like Deepseek V3 – The Decoder, accessed February 1, 2025, https://the-decoder.com/allen-ai-claims-its-new-tulu-3-405b-open-source-model-rivals-top-performers-like-deepseek-v3/
- AI World War 1 Just Began: SHOCKING New AI (Tülu 3) Destroys DeepSeek & OpenAI!, accessed February 1, 2025, https://www.youtube.com/watch?v=nJjuYTpHQEE
- en.wikipedia.org, accessed February 1, 2025, https://en.wikipedia.org/wiki/DeepSeek#:~:text=DeepSeek%20(Chinese%3A%20%E6%B7%B1%E5%BA%A6%E6%B1%82%E7%B4%A2%3B,large%20language%20models%20(LLMs).
- What is DeepSeek? — everything to know, accessed February 1, 2025, https://www.tomsguide.com/ai/what-is-deepseek-everything-to-know
- Deepseek AI Vs Open AI: A Comprehensive Comparison – EUCLEA Business School, accessed February 1, 2025, https://www.euclea-b-school.com/deepseek-ai-vs-open-ai-a-comprehensive-comparison/
- aws.amazon.com, accessed February 1, 2025, https://aws.amazon.com/blogs/aws/deepseek-r1-models-now-available-on-aws/#:~:text=This%20leads%20us%20to%20Chinese,parameters%20on%20January%2020%2C%202025.
- What Is DeepSeek, the New Chinese OpenAI Rival?, accessed February 1, 2025, https://time.com/7210296/chinese-ai-company-deepseek-stuns-american-ai-industry/
- You Won’t Believe the NEW AI Models That Beat DEEPSEEK! – YouTube, accessed February 1, 2025, https://www.youtube.com/watch?v=Nh02XzqmI20
- DeepSeek AI: Open-Source Models Revolutionizing Language, Reasoning, and Multimodal AI | Encord, accessed February 1, 2025, https://encord.com/blog/deepseek-ai/
- Ai2 Releases Tülu 3 405B The Open-Source AI Model that Challenges Industry Titans, accessed February 1, 2025, https://digialps.com/ai2-releases-tulu-3-405b-the-open-source-ai-model-that-challenges-industry-titans/
- Scaling the Tülu 3 post-training recipes to surpass the performance of DeepSeek V3 | Ai2, accessed February 1, 2025, https://allenai.org/blog/tulu-3-405B
JOIN US
Sign up to receive occasional news from Rod.