Summary:
1. DeepSeek, an AI startup, released a new model, R1-0528, which is already gaining traction in the AI community.
2. TNG Technology Consulting GmbH introduced DeepSeek-TNG R1T2 Chimera, a faster and more efficient adaptation of the original model.
3. R1T2 uses an Assembly-of-Experts method to combine the strengths of different models, resulting in a more capable and cost-effective solution.
Article:
Are you looking for smarter insights in the field of enterprise AI, data, and security? Sign up for our weekly newsletters to stay updated on the latest developments. Subscribe now to get only what matters to you.
Chinese AI startup DeepSeek, a branch of High-Flyer Capital Management, recently launched the latest version of its popular open-source model, DeepSeek R1-0528. Building on the success of its predecessor, DeepSeek-R1, which garnered attention for its affordability and performance on reasoning tasks, R1-0528 is already being embraced by developers and enterprises due to its Apache 2.0 license.
In a noteworthy development, German firm TNG Technology Consulting GmbH unveiled DeepSeek-TNG R1T2 Chimera, the newest addition to its Chimera large language model family. R1T2 offers enhanced efficiency and speed, achieving intelligence benchmark scores comparable to R1-0528 while using significantly fewer output tokens, resulting in faster inference and reduced compute costs.
Powered by TNG’s Assembly-of-Experts (AoE) technique, R1T2 integrates the strengths of DeepSeek-R1-0528, DeepSeek-R1, and DeepSeek-V3-0324 models to deliver a highly capable yet cost-effective solution for enterprise and research applications. This innovative approach allows R1T2 to maintain high reasoning capabilities while minimizing inference costs.
Differentiating itself from the Mixture-of-Experts (MoE) architecture, AoE focuses on merging weight tensors from multiple pre-trained models to create a new, more efficient model. By selectively combining expert tensors responsible for specialized reasoning, AoE enables Chimera models to inherit reasoning strength without compromising efficiency.
When it comes to performance and speed, R1T2 excels by achieving 90% to 92% of the reasoning performance of DeepSeek-R1-0528 while generating concise responses using approximately 40% fewer tokens. This reduction in output length translates to a 60% decrease in compute load, resulting in responses that are 200% faster than the original model.
R1T2’s release under the MIT License makes it open source and available for commercial applications. However, TNG advises European users to consider compliance with the EU AI Act, effective from August 2, 2025. For enterprises operating within the EU, it is essential to review and ensure adherence to the regulatory provisions outlined in the act.
In conclusion, R1T2’s introduction offers tangible benefits for enterprise technical decision-makers, including lower inference costs, high reasoning quality without overhead, open-source flexibility, and emerging modularity. As the AI landscape continues to evolve, R1T2 represents a significant advancement in efficient and cost-effective reasoning models, paving the way for future innovations in the field.
