Germany’s TNG Technology Hacks AI Speed with Supercharged DeepSeek R1T2 Variant

If a speedier, more sophisticated AI model sounds like an elusive dream, Germany’s TNG Technology just turned that into reality. A new adaptation of the DeepSeek Platform’s open-source model DeepSeek R1-0528, often hailed as an AI game-changer, is leaving a remarkable footprint on the AI and global business landscapes. Aptly named DeepSeek-TNG R1T2 Chimera, the model is the latest addition to TNG’s Chimera large language model family. Boasting up to 90% of R1-0528’s intelligence benchmark scores, R1T2 demonstrates a significant increase in efficiency and speed. However, it curiously reduces the pressure on computational resources by generating responses with less than 40% of R1-0528’s output token count. The exciting result: shorter, quick-fire responses translating to reduced inference and computational costs. TNG’s gain comes courtesy of its proprietary Assembly-of-Experts (AoE) technique used to develop large language models (LLMs) by strategically integrating the weight tensors from multiple pre-trained models. The impressive R1T2 is an example of the successful implementation of the AoE methodology. It innovatively merges three parent models: DeepSeek-R1-0528, DeepSeek-R1, and DeepSeek-V3-0324, to achieve a hybrid capable of maintaining high reasoning ability while substantially cutting down inference costs. Without the need for further fine-tuning or retraining, R1T2 inherits the strenuous reasoning capabilities of R1-0528, the structured thoughtforms of R1, and the concise, instruction-oriented behavior of V3-0324. The result is a more compact, yet versatile model prepared for both enterprise usage and research applications.