@adlrocha - LLMs are Rediscovering the Memory Hierarchy

Last week DeepSeek-V4.1-Flash went live, and as it happens whenever DeepSeek releases a new model, the AI sphere went crazy. Going from v4 to v4.1 may not seem like a major release, but according to many experts in the field, the changes in the architecture could have easily justified this model being DeepSeek-V5.With this, I couldn’t miss the opportunity to extend the series of posts that I started with Kimi K3, the post-training piece and the Flash pair, and dedicate a post explaining what the changes in the architecture of this DeepSeek model means for the field. Once more, this model prioritises performance improvements and is designed in a way that roughly a third of this model doesn’t need to be hosted in the GPU.Some of the tricks introduced in this new DeepSeek architecture could already be seen in the Qwen3.8-Flash-Next architecture I described in this post, which is interesting because, and this is going to become my most repeated sentence this year, “it seems like we are…