Unverified50% confidenceBenchmarkExact time
单卡RTX 4090跑4Bit量化的通义千问3-27B约每秒生成49个字
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/27/2026
First Seen
Valid until: 11/25/2026
Sources
Related Entities
Related Claims
Unverified128K长度的序列所需的注意力计算量是4K长度序列的1024倍72% similarUnverified将1536维向量分解为16个96维子空间、每个子空间使用256个码字,则整个向量仅需16字节表示,压缩比高达384:1,每次距离计算只需16次查表操作70% similarUnverifiedSymbol实现了单仓库平均250毫秒索引、单次查询仅1.5毫秒的性能表现70% similarUnverified在配备96GB显存的M2 Max上以1-bit精度运行时,简单任务生成速度稳定在约28-30 tokens/秒68% similarVerified9600波特率下UART每秒可传输约960个字节(每字节需10个比特:1起始位+8数据位+1停止位)68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/810112API
curl https://kongchang.com/api/v1/knowledge/claims/810112MCP
get_claim(id=810112)