A discussion on X critiques HBM capacity as an overblown bottleneck in AI inference, arguing that recent optimizations like V4.1 Flash have reduced memory requirements per token from 3.5KB to 890 bytes through techniques such as CSA2 and SWA Bounded Replay. The author contends that bandwidth, not capacity, is the true limiting factor in decode operations, and that moving from 12-Hi to 8-Hi HBM stacks reflects this reality rather than supply constraints.
NVIDIA Research demonstrates accelerated video generation using MiniMax-H3 on a single DGX Spark, producing 768p video in one minute through a two-stage pipeline that generates a 384p draft and refines it to higher resolution, with detailed prompts for stop-motion and photorealistic video examples.
DeepSeek released the V4.1 Flash model, a 552B parameter MoE model with native multimodal vision capabilities that outperforms V4 Pro across benchmarks while significantly reducing API pricing. The model features a new Causal-Encoder-Decoder architecture with dramatically reduced KV Cache requirements and improved inference speed, with V4 Pro being phased out in favor of the new model.