Doesn't evict live KV: Minima-KV compresses long-context cache 3.5x with mixed-format attention
Long-context serving is throttled by the KV cache, and eviction is the usual remedy. Minima-KV instead keeps recent Anchor pages in FP8 while shoving older pages into packed TQ3, merging the two through a globally normalized online-softmax so no cache-sized shadow is needed. The paper reports 3.5x compression versus BF16, no quality loss on RULER needle tests, and sub-1-point deltas over LongBench v2. Every live-request page stays addressable, the part that makes it deployment-ready.