TNG Technology Consulting has unveiled the DeepSeek-Tng R1T2 Chimera, a new Expert Assembly (AOE) model that brings intelligence and speed together through innovative model merging strategies. Built from three high-performance parent models – R1-0528, R1 and V3-0324-R1T2.
Expert assembly: Large-scale effective model composition
Traditional LLM training and fine-tuning require a lot of computing resources. TNG solves this with its Installation Engineering (AOE) approach and incorporates large-scale Experts (MOE) models at the weight tensor level without retraining. This strategy can enable linear time to build new models, thus inheriting the abilities of multiple parents. The architecture of R1T2 combines R1’s expert tensor with the basis of V3-0324 and selectively includes improvements to R1-0528, optimizing the trade-off between inference cost and inference quality.
Speed growth and intelligence tradeoffs
In the benchmark comparison, R1T2 is 20% faster than R1, while R1-0528 is more than twice as fast. These performance growths are largely attributed to the decrease in their output token length and selective expert tensor integration. While it has a slight R1-0528 in RAW Intelligence, it outperforms R1 significantly in advanced benchmarks such as GPQA Diamond and Aime-2024/2025.
Furthermore, the model retains the…n inference trace, which only occurs when R1’s contribution to the merger crosses a specific threshold. This behavioral consistency is crucial for the application of reasoning that requires a gradual reflection.
Emerging properties in parameter space
R1T2 confirms the finding from the accompanying research paper that model merging can produce feasible models throughout the interpolation space. Interestingly, the smart properties gradually change, but behavioral markers (e.g. consistent use) suddenly appear at 50% R1 weight ratio. This suggests that certain features are located in different subspaces of the LLM weight landscape.
By merging only the routing expert tensors and leaving behind other components from V3-0324 complete (e.g., note, focus and share MLP), R1T2 maintains a high inference score while avoiding verboseness. This design leads to what TNG calls “consistency of thought”, a behavioral characteristic in which reasoning is not only accurate but also concise.
Early discussions in the Reddit Locallama community highlight the actual impression of R1T2. Users praised the model for its responsiveness, token efficiency, and a balance between speed and coherence. “This is the first time the Chimera model feels like a real upgrade of speed and quality,” one user noted. Another noted that it performs better in mathematically heavy contexts than the previous R1 variant.
Some red diseases have also been observed that R1T2 exhibits a more grounded role, avoiding hallucinations that are more stable than models based on R1 or V3. This urgent feature is especially important for developers looking for a stable LLM backend for production environments.
Open weights and availability
R1T2 is publicly obtained under MIT license on Embracing Faces: DeepSeek-Tng R1T2 Chimera. This version encourages community experiments, including downstream fine-tuning and enhanced learning. According to TNG, on-premises through Chutes serverless inference platform has processed nearly 5 billion tokens per day.
in conclusion
DeepSeek-TNG R1T2 Chimera demonstrates the potential of Experts construction to generate performance, effective LLM without gradient-based training. By strategically combining R1’s inference capabilities, V3-0324’s token design design, and R1-0528’s enhancement function, R1T2 has established a new standard for balanced model design. Its open release under the MIT license ensures accessibility, making it a powerful candidate for developers looking for fast, capable and customizable large language models.
Since model merging proves feasible even on the 671B parameter scale, TNG’s R1T2 can serve as a blueprint for future experiments with parameter space interpolation, enabling more modular and interpretable LLM development.
Check Hug the paper on the face and open the weight. All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Asif Razzaq is CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, ASIF is committed to harnessing the potential of artificial intelligence to achieve social benefits. His recent effort is to launch Marktechpost, an artificial intelligence media platform that has an in-depth coverage of machine learning and deep learning news that can sound both technically, both through technical voices and be understood by a wide audience. The platform has over 2 million views per month, demonstrating its popularity among its audience.
