DeepSeek-TNG R1T2 Chimera Boasts 200% Speed Increase Over R1-0528 with Enhanced Reasoning

TNG Technology Consulting GmbH has announced the release of its new artificial intelligence model, DeepSeek-TNG R1T2 Chimera, now available on Hugging Face under the MIT license. This "Tri-Mind Assembly-of-Experts" model is built from three parent models: DeepSeek-R1-0528, DeepSeek-R1, and DeepSeek-V3-0324. The company positions it as a significant advancement, operating at a "sweet spot in intelligence vs. output token length."

The R1T2 Chimera demonstrates notable performance improvements over its predecessors. According to TNG Technology Consulting's announcement, the new model is "about 20% faster than R1, and more than twice as fast as R1-0528." This increased speed is primarily attributed to its ability to generate responses with significantly fewer output tokens, leading to faster inference and reduced computational costs.

In terms of intelligence, the model shows a substantial leap, being "significantly more intelligent than R1 in benchmarks such as GPQA Diamond and AIME-24/25." While it may not quite match the highest intelligence level of R1-0528, it marks a major improvement over the first R1T Chimera, particularly in achieving "think-token consistent" behavior. This consistency ensures more compact and orderly reasoning processes.

The DeepSeek-TNG R1T2 Chimera leverages an Assembly-of-Experts (AoE) method, a technique that merges weight tensors from multiple pre-trained models rather than traditional fine-tuning or retraining. This approach allows the model to inherit the reasoning strength of R1-0528, the structured thought patterns of R1, and the concise output characteristics of V3-0324. Its open-source nature under the MIT license makes it accessible for both research and commercial applications.

Despite its advancements, the R1T2 Chimera currently has limitations, notably in its support for function calling, a feature inherited from its DeepSeek-R1 lineage. TNG Technology Consulting also advises users in the European Union to assess compliance with the upcoming EU AI Act, which takes effect on August 2, 2025. The company expressed anticipation for community experiments and feedback on the new model.