Which statement most accurately captures why GQA, not MQA, became the production default?

**Multi-head attention** gives every query head its own key and value head. That is great for quality, because each head can specialize, but it means the KV cache stores keys and values for all of those heads. At long context that cache becomes the dominant memory cost during decoding. Multi-query attention attacks that directly: keep all the query heads, but let them share a single key head and a single value head. The cache shrinks dramatically. The catch is that collapsing to one shared head removes a lot of the specialization, and quality drops on harder tasks. Grouped-query attention is the compromise. Instead of one shared key and value head or a full set, you pick a small number of KV heads, and each one is shared by a group of query heads. A group size of eight is the common choice. The correct option captures exactly this: GQA hits roughly MHA-level quality at a tunable group size while still delivering the KV-cache reduction. It does not win on compute or parameter count, which is why the other options are wrong.

**Grouped-query attention** sits between two extremes. Multi-head attention gives every query head its own key and value head. Multi-query attention shares one key and one value head across all query heads. GQA picks an intermediate number of KV heads. Each KV head is shared by a group of query heads. The group size is a single integer knob. At one KV head it becomes MQA. At a full set it becomes MHA. Teams settled near eight KV heads in practice. The payoff is on the memory axis, not compute. The KV cache shrinks by the ratio of query heads to KV heads. That cuts the bytes streamed from HBM on every decode step. The query times key matmul still runs for every query head, so compute barely changes. MQA delivers the same cache savings but shows measurable quality regressions on harder tasks. GQA recovers most of that lost quality at a small memory cost. That is why it became the default in Llama-style models. The wrong options confuse the axis. It is not fewer dot products. It is not fewer query parameters. MQA also produces a smaller cache than MHA, not larger. The real story is tunable cache reduction with quality held near MHA.