Unverified50% confidenceFactExact time
The architecture's sliding window is 128 tokens, and the KV cache of the window layers stops growing after 128 positions, no matter how long the context is.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe sliding window with bounded replay only replays the most recent 128 tokens, accepting approximation errors to solve the cache offloading problem.74% similarVerifiedModel weights cannot fill up VRAM entirely, as space must be reserved for the KV cache and inference processes; the KV cache required for a 128K context can reach as high as several GB or even exceed the model weights themselves.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1061807API
curl https://kongchang.com/api/v1/knowledge/claims/1061807MCP
get_claim(id=1061807)