DeepSeek V4 reworks transformer architecture

My read is that the post is strongest when it focuses on architectural bets rather than local-runnable hype.

• –Hybrid attention is interesting because it keeps the stack attention-native instead of swapping in a different sequence model family.
• –The residual-path redesign is the most consequential claim; if stable in practice, it could matter more than another incremental attention tweak.
• –The hardware comments are directionally right: most users will consume this through hosted endpoints or distilled variants, not full local inference.
• –Some claims are still community interpretation rather than hard, reproducible evidence, so the post works best as a discussion starter rather than a definitive technical summary.