Unverified60% confidenceBenchmarkExact time
在 DeepSeek R1 670B（MXFP4 精度）测试中，Jalapeño 每瓦吞吐约 1.7 倍（19,641 对 GB300 的 11,781），端到端延迟从 5.99 秒降到 1.65 秒
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
9/24/2026
First Seen
Valid until: 12/23/2026
Sources
Related Entities
Related Claims
Unverified在GSM8K上OQ4E用时2450.2秒，8bit高达5471.2秒，耗时差了两倍多66% similarUnverified13B-14B级别模型（4-bit量化）在该配置下速度下降到每秒2-5个token66% similarUnverifiedDeepSeek蒸馏版参数量从1.5B到70B不等65% similarUnverified实测在RTX 3060（6GB显存）、32GB RAM上运行预量化Krea2 Turbo，1024×1024分辨率8步Euler采样，速度为1.78 s/it，总耗时17.64秒64% similarUnverifiedJalapeño额定功耗700W，实测持续功耗未超过550W；对照GB200额定1200W，GB300额定1400W64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945835API
curl https://kongchang.com/api/v1/knowledge/claims/945835MCP
get_claim(id=945835)