Unverified50% confidenceBenchmarkExact time
Nvidia CentML engineer Benjamin, through upstream vLLM optimization, enabled the 2.8-trillion-parameter Kimi K3 to achieve an inference speed of 423 tok/s/user on GB300 NVL72, a performance improvement of 1.68 times.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/19/2026
First Seen
Valid until: 11/2/2026
Sources
Related Entities
Related Claims
UnverifiedOn the B300 benchmark, Kimi K3's new optimization solution on vLLM achieved a 2.2 to 2.8× throughput improvement compared to version v0.27.1.71% similarUnverifiedAccording to the documentation, Kimi K2.8 performs at a level comparable to K3 but has more efficient reasoning capabilities, supporting a 1-million-token context window, image and video input, and three reasoning modes: low, high, and max.68% similarUnverifiedKimi K3使用自研的Kimi Delta Attention机制，解码速度最高提升6.3倍，上下文窗口支持100万token64% similarUnverifiedMixtral-8x7B achieves performance close to that of a 70B dense model with a total of 47B parameters, while offering faster inference speed64% similarUnverifiedIn actual testing, the Top Speed configuration completed the run in 14 minutes consuming 2 million tokens; the Kimi K3 model in the Open Weights configuration failed entirely and could not output correct JSON; and the planning phase of Opus 5 in the Frontier configuration ran for 13 minutes.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065048API
curl https://kongchang.com/api/v1/knowledge/claims/1065048MCP
get_claim(id=1065048)