Inty News
@__Inty__
ๆฌข่ฟ่ฟๆ็่ๅคฉ็พค ๐
๊ฐ์
August 2018
53 ํ๋ก์ ์ค 750.1K ํฌ
1 token/็ง
๐
You can Run DeepSeek V4 Flash (33B MoE) running on 8 GB RAM, CPU-only, zero GPU.
- Cold first-token: 5.33s
- Max RSS: 5.9-6.2 GiB
- Process swaps: 0
The entire model stayed memory-mapped on NVMe. Linux demand-paged only the needed weights into RAM.
This is not fast, Itโs not production.
But it proves something important: Model size vs RAM is now a performance problem, not a hard barrier.
VRAM, RAM, NVMe is becoming a real memory hierarchy for local AI.
-
๋ ๋ณด๊ธฐ
