DeepSeek V4 Just Made 1-Million-Token Context Look Cheap — Here’s the Trick

An open-source 1.6T MoE that runs at 10% of the KV cache and a tenth of GPT-5’s price. The architecture is the story.