Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

End-to-end Time-to-First-Token (TTFT) on Long-context sequences

257TTFT (ms)

FlashPrefill

-1,882.812,560.8527,004.541,448.15Mar 6, 2026
Updated 1mo ago

Evaluation Results

MethodLinks
2026.03
2571.04
2026.03
267-
2026.03
4061.04
2026.03
423-
2026.03
4471.05
2026.03
471-
2026.03
5191.18
2026.03
613-
2026.03
8511.07
2026.03
912-
2026.03
9131.07
2026.03
976-
2026.03
1,0901.47
2026.03
1,603-
2026.03
1,7071.2
2026.03
1,8601.18
2026.03
2,048-
2026.03
2,189-
2026.03
2,3311.98
2026.03
3,5341.34
2026.03
3,6181.49
2026.03
4,635-
2026.03
4,735-
2026.03
4,9093.1
2026.03
5,398-
2026.03
7,3231.72
2026.03
7,4811.97
2026.03
10,7025.02
2026.03
12,563-
2026.03
14,774-
2026.03
15,0613.02
2026.03
15,203-
2026.03
15,2132.45
2026.03
37,239-
2026.03
45,464-
2026.03
53,752-