Unverified50% confidenceBenchmarkExact time
On a single H100 with torch.compile, BF16 precision and two 256x256 RGB inputs, REKI produces a complete action chunk in only 21.8 milliseconds, 35.5% faster than π0.5 (33.8 ms).
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 a single H100 in warm state, the standard version (32B, 4-step) takes 843 milliseconds to generate and decode a 22-frame 540p video.71% similarUnverified在M4 Max上进行FP16精度测试，hy3d shape（small）任务运行时间20.9秒，峰值内存约5.6GB67% similarUnverifiedRunning on 4 DGX Spark units with VLLM's Tensor Parallel 4, measured decoding speeds were approximately 70 tokens/second for code and HTML, approximately 45 for JSON, and approximately 30 for prose and planning tasks, while DeepSeek's own API can exceed 300 tokens/second.66% similarUnverifiedUnder a standard synthetic workload of 1k prompt / 512 completion, V4.1 Flash 1017 has a TTFT of 185 milliseconds, total throughput of 480 tokens/s, and KV cache efficiency of 0.65x memory.66% similarUnverifiedThe UP host tested MiniCPM-2B on a computer with 4G VRAM and achieved a generation speed of 57 tokens per second, with a first-token latency of only 0.22 seconds.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1070502API
curl https://kongchang.com/api/v1/knowledge/claims/1070502MCP
get_claim(id=1070502)