Unverified50% confidenceBenchmarkExact time
在80亿参数模型的文本生成测试中，M5 Ultra每秒188 tokens，M3 Ultra每秒129 tokens，仅为1.45倍
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
UnverifiedM3 Ultra运行四位DeepSeek R1时，1000 token上下文约每秒生成19个token占用392GB，16000 token时降到约每秒15个占用410GB78% similarUnverified在AmorphousDiskMark测试中M5 Ultra顺序读取达14,888MB/s、写入接近20,000MB/s，M3 Ultra读取约6500MB/s、写入2700MB/s76% similarUnverified在4×RTX 3090 SXM4配置下，GLM-5.2量化版本的Token生成速度约为6.3-6.8 tokens/秒，提示评估速度约为10.6 tokens/秒75% similarUnverified该配置下35B模型实测单人最高速度约80 Tokens/秒，并发可达500 Tokens/秒以上72% similarUnverified用2840亿参数的DeepSeek V4 Flash实测，token生成从37提升到53 tokens/s（约1.5倍），Prompt Processing从483跃升到1485 tokens/s（约三倍）72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/944991API
curl https://kongchang.com/api/v1/knowledge/claims/944991MCP
get_claim(id=944991)