Unverified50% confidenceBenchmarkExact time
Across microbenchmarks of five shapes, v3 (split-K + autotuned tile) is not faster than v2 (fused kernel) in any case, and v2 is 1.0 to 1.4 times faster than v3.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/12/2026
First Seen
Valid until: 10/26/2026
Sources
Related Entities
Related Claims
UnverifiedThe Q3_XXS model paired with an F16 KV cache scores approximately 80% on the needle test, while Q3_K_XL paired with a kvarn3 KV cache scores close to 100%.62% similarUnverifiedV4.1-Flash scored 74.2 points on a certain programming test dimension, an improvement of nearly 20 points over V4-Flash.60% similarUnverifiedV4 Pro's parameter count is nearly 6 times that of the Flash model, yet its actual performance has not opened up a significant gap.59% similarUnverifiedVTracer was chosen because of its compact size (approximately 0.66 MiB) and because it produced the best vectorization results in baseline testing.59% similarUnverifiedThe three-level cache improved generation speed under Q4 weights from 1.83 tokens per second to 8.89, an improvement of nearly 5x.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1006838API
curl https://kongchang.com/api/v1/knowledge/claims/1006838MCP
get_claim(id=1006838)