Unverified50% confidenceFactExact time
v3.36.0 adds usage counting for 30-minute cache writes (PR #4111) and compaction trigger IDs (PR #4122).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/21/2026
First Seen
Valid until: 1/19/2027
Sources
Related Entities
Related Claims
UnverifiedFrom V1 to V4.1 Flash, the single-token cache was compressed from 389,000 bytes to 890 bytes, a total compression of 437 times.64% similarUnverifiedReplacing the text encoder from int8_convrot (1.67 s) with qwen3vl_32b_minimax_h3_nvfp4_awq (0.7 s) saves nearly one second per generation.64% similarUnverifiedVPU latency is about 0.32 ms/step, with throughput exceeding 3000 FPS.64% similarUnverifiedOn a single NVIDIA A40 with fp32 and TF32 disabled, batch size 1 and greedy decoding, after GPT-2 small uses the KV cache, single-token latency at context length 1000 drops from 29.69 ms to 4.95 ms, about 6x.63% similarUnverifiedIn measurements at dim=256 and dim=512, VPU's average latency is about 0.32ms in both cases, with equivalent FPS of about 3088-3099.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1081990API
curl https://kongchang.com/api/v1/knowledge/claims/1081990MCP
get_claim(id=1081990)