TNG Tech Unveils DeepSeek-TNG R1T2 Chimera, a 200% Faster AI Model

A little over a month ago, Chinese AI startup DeepSeek released the latest version of its open-source model, DeepSeek R1-0528. This release followed the initial success of DeepSeek-R1, which garnered significant attention for its low training costs and high performance on reasoning tasks. Now, TNG Technology Consulting GmbH, a 24-year-old German firm, has taken this model to the next level with the introduction of DeepSeek-TNG R1T2 Chimera.

What Changed Technically?

• Assembly-of-Experts (AoE) Method: TNG's new model leverages an innovative technique called Assembly-of-Experts (AoE), detailed in a paper published on arXiv. This method involves selectively merging the weight tensors (the internal parameters of neural networks) from multiple pre-trained models.
• Tri-Mind Configuration: R1T2 introduces a "Tri-Mind" configuration, combining three parent models: DeepSeek-R1-0528, DeepSeek-R1, and DeepSeek-V3-0324. This integration aims to maintain high reasoning capabilities while significantly reducing inference costs.

Key Performance Metrics

• Efficiency and Speed: R1T2 delivers a notable boost in efficiency and speed:
  • Achieves around 90% of R1-0528’s intelligence benchmark scores.
  • Generates responses with less than 40% of R1-0528’s output token count, leading to faster inference and lower compute costs.
  • Approximately 20% faster than the regular R1 (released in January) and more than twice as fast as R1-0528 (the May update from DeepSeek).

Community Response

The AI developer community has responded enthusiastically to this new model. Vaibhav (VB) Srivastav, a senior leader at Hugging Face, praised the development on X:

DAMN! DeepSeek R1T2 – 200% faster than R1-0528 & 20% faster than R1. Significantly better than R1 on GPQA & AIME 24, made via Assembly of Experts with DS V3, R1 & R1-0528, and it’s MIT-licensed, available on Hugging Face.

Technical Details

• Model Architecture: The Tri-Mind configuration in R1T2 is designed to balance the strengths of its parent models. By merging the weight tensors from DeepSeek-R1-0528, DeepSeek-R1, and DeepSeek-V3-0324, TNG has created a model that excels in reasoning tasks while being highly efficient.
• Benchmarks: R1T2’s performance on intelligence benchmarks is impressive, maintaining high accuracy with significantly reduced computational requirements. This makes it an attractive choice for developers looking to deploy AI models at scale without breaking the bank.

Implications for Practitioners

For practitioners and businesses, the release of DeepSeek-TNG R1T2 Chimera means:

• Cost Savings: Lower compute costs due to faster inference times.
• Improved Efficiency: Shorter response times can enhance user experience in applications like chatbots and virtual assistants.
• Flexibility: The MIT license allows for broad adoption and customization, making it a versatile tool for various AI projects.

Conclusion

TNG Technology Consulting GmbH's DeepSeek-TNG R1T2 Chimera represents a significant advancement in the field of large language models. By combining the strengths of multiple pre-trained models through the Assembly-of-Experts method, TNG has created a model that is both efficient and powerful. This development underscores the ongoing innovation in open-source AI and highlights the potential for further advancements as more developers contribute to and build upon these models.