deepseek-v4.1POST

DeepSeek V4.1 Flash on 4×DGX Spark: 97.7 tok/s single request, 16-concurrency throughput drops 38%

Native Mia/SGLang versus TensorFold EXL3 3.5bpw over dual RoCE: single-request throughput rises from 79.2 to 97.7 tok/s, 16-concurrency falls from 434.7 to 267.6, and 128K first-token latency goes from 26.9 to 55.3 seconds.

Devye

@XDevyeArchived 12 days agoOriginally posted 2026.10.19

Devye @XDevyeOriginally posted 2026.10.19

Open original on X

4×DGX Spark 实测 DeepSeek V4.1 Flash 1017： Mia/SGLang 原生 → TensorFold EXL3 3.5bpw 双 RoCE 端到端吞吐（三轮均值）：单请求 79.2→97.7 tok/s（+23%） 16并发 434.7→267.6（-38%） 128K首token：26.9→55.3秒。

x.com/xdevye/status/2112009853363490927Archived 2026.10.19 12:04