WHAT NEW IN X11.01 SUN

AI collections from X · 22 collections · 64,204 posts

deepseek-v4.1IMAGE

DeepSeek V4.1 Flash 748B hits nearly 20 tok/s on a single 48GB modded 4090

A hands-on test runs a quantized V4.1 Flash on a 48GB modded 4090 with 128GB RAM, caching hot experts in VRAM to lift throughput from 5–6 to nearly 20 tok/s.

洋仔实验室
@yang3izcArchived 10 days agoOriginally posted 2026.10.21
DeepSeek V4.1 Flash 748B hits nearly 20 tok/s on a single 48GB modded 4090
洋仔实验室 @yang3izcOriginally posted 2026.10.21
Open original on X

748B 的模型,单张 4090 真跑起来了。 黄总实测从 5–6 tok/s 压到近 20 tok/s。 关键在于:把热专家塞进显存当缓存。 硬件前提是 48G 魔改 4090 + 128G 内存,不是普通卡。 想试量化版,模型页在:https://huggingface.co/Baekpica/DeepSeek-V4.1-Flash-1017-VQ-GGUF 我好奇的是:普通 24G 4090 能跑出多少?

x.com/yang3izc/status/2112876684257157406Archived 2026.10.22 12:01