Recent LLM Architecture Changes: From Gemma 4 to DeepSeek V4

LLM近期重大架构进化一览：从Gemma 4到DeepSeek V4

Jiqizhixin translated Sebastian Raschka’s blog on recent LLM architecture changes, covering long-context cost reductions in Gemma 4, Laguna XS.2, and ZAYA1-8B; the article states that Gemma 4 E2B saves about 2.7GB of KV cache at 128K context with bfloat16 precision.

Why it matters: HKR-H/K/R pass: notable model names, a concrete 128K bf16 KV-cache saving, and inference-cost relevance. As a translated survey rather than a release, it stays in the 72–77 featured band.