Gemma 4 Draws KV Cache Complaints

Gemma 4 is a strong launch on paper, but this thread shows how quickly “best open model” claims collide with real deployment math. If the cache footprint forces aggressive quantization just to stay under budget, a lot of local users will pick the more memory-efficient model instead.

• –Google positions Gemma 4 as a four-size family with E2B, E4B, 26B MoE, and 31B dense variants, plus up to 256K context and strong benchmark results.
• –The complaint here is not about raw quality; it’s about VRAM economics, with users saying 40GB still isn’t enough for a Q8 31B setup at modest context without KV quantization.
• –That creates a practical head-to-head with Qwen3.5-27B, which commenters report fits more comfortably at full context and is already viewed as a safer local default.
• –For local inference, cache efficiency matters as much as benchmark rank. A model that wins benchmarks but loses on memory footprint can still lose adoption on consumer hardware.
• –The launch is still relevant: Gemma 4 is clearly aimed at developer workstations and agentic workloads, but serving stacks will need to keep improving cache handling to make the promise feel real.