Unverified50% confidenceBenchmarkExact time
The M3 Ultra begins to degrade at around 16K tokens, and by 500K tokens its incremental speed drops to one-third to one-quarter of its peak, whereas the Spark maintains about one-half at the same point.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
Unverified13B–14B class models (4-bit quantization) see speeds drop to 2–5 tokens per second under this configuration61% similarUnverifiedQ4_K_M量化文件大小17.77GB，pp2048速度96.8 tok/s，tg128速度11.4 tok/s61% similarUnverifiedCollapse推理速度：CPU上单条10-token上下文约0.23毫秒，MPS在batch 1024下约230万token/秒，最近邻查询约0.48毫秒59% similarUnverifiedThe reference model Supra-50M has 51.8M parameters, is 103.6 MB in bf16, 56.2 MB after 8-bit quantization, perplexity degrades from 165 to 193 when reduced to int4, and collapses entirely when compressed to ternary.59% similarUnverifiedWhen generating a 5-second 768P video with Fast Cache enabled, the M3 Ultra took 37 minutes and 14 seconds, while the M5 Ultra took 9 minutes and 24 seconds, nearly 4 times faster.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945148API
curl https://kongchang.com/api/v1/knowledge/claims/945148MCP
get_claim(id=945148)