The Snapdragon 8 Elite Extreme Gen 6 can run models with 30,000 million parameters, but there is a trick

Qualcomm today announced its new flagship, the Snapdragon 8 Elite Extreme Gen 6, about which we tell you everything in detail here. This new SoC is so powerful that, according to Qualcomm, it can run locally artificial intelligence models with 30 billion parameters. This is considerable; In fact, today we have models with about 30 billion parameters such as Gemma 4 or Qwen 3.6, which with this volume reach 80% of the precision achieved by large models in the cloud with billions of parameters.

Said like this, it sounds spectacular, however, there is a trick. The key is in the MoE model architecture. These models divide certain specialized layers and use a selection mechanism to decide which ones are involved when processing each token. This way you don’t have to load and activate the entire model, just a part of it, making it much faster and more efficient.

An MoE of 30 billion parameters can activate around 3 billion parameters routed per token. This reduces the computation relative to using the entire network at each step, but it does not eliminate the need to store its parameters nor does it guarantee a proportional speedup.

During the event we were able to see it in operation, reaching 30 Tokens per second.

This article has been possible thanks to GEEKNETIC journalists attending a presentation event in person organized by Qualcomm, which covered the travel expenses. Beyond providing information, conferences and interviews with its spokespersons, Qualcomm was not part of the preparation of the article and the content has been written, edited and published in an impartial and independent manner.