New research paper on the "Assembly of Experts" for TNG's Chimera model

Our new paper “Assembly of Experts: Linear-time construction of the Chimera LLM variants with emergent and adaptable behaviors” has been published on arXiv and Hugging Face.

The paper explains how we constructed the 671B R1T Chimera child model from the great DeepSeek AI V3-0324 and R1 parent models (谢谢!) in less than one hour of CPU time.

The Chimera model is our default reasoning model internally at TNG.

Some stats: the Chimera research prototype is currently the fifth most-popular LLM on chutes.ai with over three billion tokens per day, close to one billion of which are coming from OpenRouter. More than 160 billion tokens have been processed since release on April 26th.

If you are wondering how you can make use of the model or the assembly method, you can reach out to us.