Unverified50% confidenceBenchmarkExact time
NVIDIA's CentML team achieved 423 tokens per second per user on the 2.8-trillion-parameter Kimi K3 large model, a 1.68x speedup on the same architecture, thanks to a decoding kernel written in CUDA DSL and Mega Kernel Fusion technology, with about 12,000 lines of optimized code.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/29/2026
First Seen
Valid until: 1/27/2027
Sources
AI Buzz Weekly: GPT-6.5 Leak, New Claude Model, and MiniMax's Cost-Effectiveness
bilibiliTHE_OREGANO10/29/2026
Related Entities
Related Claims
UnverifiedRunning Kimi K2.5 on an NVIDIA B200 with real coding agent trajectories (SWE-smith, EvalScope evaluation), TokenSpeed outperforms TensorRT-LLM across the entire latency/throughput curve; at batch size 1 it is about 9% faster, and at about 100 tokens/second per user its throughput is about 11% higher.68% similarUnverifiedMegatron-LM实验数据表明，NVLink互联下8卡张量并行可达7.5×以上线性加速，而PCIe 3.0 x16环境下7B模型通常仅达2.5-3×，超过4卡后通信开销完全抵消额外算力67% similarUnverifiedWhen speculative decoding is configured properly, token generation speed can improve by about 1.5x (according to vLLM); Omer personally tested Qwen3 35B MoE with a DFlash drafter in his home lab and measured a 2x to 3x improvement, close to 3x in coding scenarios and about 1.5x in browser automation.66% similarUnverifiedDeepSeek-V4.1-Flash-1017 supports a 1 million token context, with a recommended maximum output of 256K or more.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1111905API
curl https://kongchang.com/api/v1/knowledge/claims/1111905MCP
get_claim(id=1111905)