KV Cache Study Native NIAH - 2K: leaderboard
KV Cache Study Native NIAH - 2K: Answer-prefix accuracy over 192 queries (%). KV cache updates arXiv 2611.14051v1: native Source and Target use each checkpoint for both prefix and continuation; cross-cache Reuse/Mapper are not native model observations. Six NIAH prefix lengths each have eight histories x24 queries=192. Native unmodified GLM-5.2/5.3, DeepSeek V4.1 Flash/1017, DeepSeek V4 Flash/0731 and Qwen3-30B-A3B/Instruct-2507 remain distinct checkpoints. Free generation greedy. Table 8 is the separate 585-example BABILong/QA cohort. No base/instruct merge and no reuse accuracy attributed to native model.
Metric: Answer-prefix accuracy over 192 queries (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | 100 | #58 |
| 2 | DeepSeek V4.1 Flash (1017) | 100 | #38 |
| 3 | GLM-5.3 | 99.48 | #62 |
| 4 | DeepSeek V4 Flash | 98.96 | #254 |
| 5 | GLM-5.2 | 96.88 | #137 |
| 6 | DeepSeek V4 Flash (0731) | 92.19 | #136 |
| 7 | Qwen 3 30B A3B 2507 Instruct | 76.04 | #1078 |
| 8 | Qwen 3 30B A3B | 0 | #1124 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=kv-cache-study-native-niah-2k · How It Works · Data refreshed daily, snapshot 2026-11-12.