Share your thoughts, 1 month free Claude Pro on us
See more
Home
/
Benchmarks
Long-context Retrieval on Needle-in-a-Haystack 64K-token context
Loading...
98.7
Success Rate
Full KV Cache
-0.828
25.011
50.85
76.689
Jun 29, 2026
Jun 30, 2026
Jul 1, 2026
Jul 3, 2026
Jul 4, 2026
Jul 5, 2026
Jul 7, 2026
Success Rate
TTFT (s)
Updated 17d ago
Evaluation Results
Method
Method
Links
Success Rate
TTFT (s)
Full KV Cache
Context Length=64K
2026.07
98.7
-
DepthWeave-KV
Context Length=64K
2026.07
96.1
-
TailorKV
Context Length=64K
2026.07
92.6
-
ChunkKV
Context Length=64K
2026.07
91.2
-
Eigen Attention
Context Length=64K
2026.07
90.1
-
MiniCache
Context Length=64K
2026.07
89.4
-
KVSharer
Context Length=64K
2026.07
88.7
-
PyramidKV
Context Length=64K
2026.07
88.1
-
SnapKV
Context Length=64K
2026.07
86.5
-
H2O
Context Length=64K
2026.07
83.8
-
StreamingLLM
Context Length=64K
2026.07
79.6
-
Qwen3-30B-A3B-Instruct-2507-FP8 with HGA group-level routing
Needle depth=Overall,...
2026.06
3
-
Qwen3-30B-A3B-Instruct-2507-FP8 with HGA group-level routing
Needle depth=25%, fine...
2026.06
-
948
Qwen3-30B-A3B-Instruct-2507-FP8 with HGA group-level routing
Needle depth=50%, fine...
2026.06
-
1,053
Qwen3-30B-A3B-Instruct-2507-FP8 with HGA group-level routing
Needle depth=75%, fine...
2026.06
-
1,332
Feedback
Search any
task
Search any
task