Deepseek just did the impossible

DVDeepSeek V4.1 FlashDeepSeekAvailable in EU

Open-source Flash model from DeepSeek matching frontier performance with a KV cache about 437x smaller than V1 (890 bytes per token). It uses a split encoder/decoder design, CSA2 cache sharing, a hierarchical sparse indexer, SWA bounded replay and single-pass mHC. Ranks first among open models on LiveBench and first on Vals Index, is the cheapest, and runs over 200 tokens/sec via API.

LALuma AgentsLuma AILimited in EU

Creative AI agent that accesses top image and video models and autonomously runs whole workflows such as brand kits and marketing assets. Reusable Luma skills can turn a product photo into UGC videos or a 360° orbit view. This was the video's sponsor segment.

ENEngramDeepSeekUnverified

A separate 168B-parameter memory module that sits in server RAM rather than GPU memory and stores static facts. This frees GPU memory for reasoning.

Supplementary component that lets the model output multiple tokens at a time instead of one, speeding up generation. The creator covered it in a separate explainer video.