以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

FlashMemory-DeepSeek-V4

4 atoms · 跨 1 天 · 首见 2026-06-13 · 最近 2026-06-13

三色: 🟦 fact 4 · 🟥 take 0 · stance ▲0/▼0/◆4

来源: X 4

时态: fresh:4

标签: 好数字:4

2026-06-13 · KV cache reduction percentage via Lookahead Sparse Attention at 500K context
Lookahead Sparse Attention cuts the KV cache by over 90% at 500K context
tags: 好数字 · → daily
value: qty=over 90% · direction=reduction
2026-06-13 · compressed KV cache size relative to full size
compressing it to just 13.5% of full size
tags: 好数字 · → daily
value: qty=13.5%
2026-06-13 · accuracy maintained or improved on RULER, LongBench-v2, and LongMemEval
maintaining or improving accuracy on RULER, LongBench-v2, and LongMemEval
tags: 好数字 · → daily
2026-06-13 · resident KV cache fraction on GPU via CSA sparsification
sparsifying DeepSeek-V4's CSA KV-cache on the fly to keep only ~10–15% resident on GPU
tags: 好数字 · → daily
value: qty=~10–15%

2026-06-13 · fact · KV cache reduction percentage via Lookahead Sparse Attention · → daily
2026-06-13 · fact · compressed KV cache size relative to full size · → daily
2026-06-13 · fact · accuracy maintained or improved on RULER, LongBench-v2, and · → daily
2026-06-13 · fact · resident KV cache fraction on GPU via CSA sparsification · → daily