Md Ismail Šojal 🕷️(@0x0SojalSec):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, It’s 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. -

Md Ismail Šojal 🕷️

@0x0SojalSec

Cyber_Security_Re-searcher || Ai Re-searcher || AI-Sec|| Malware Analysis II iOS || Pwn || 0SINT || Project AI-StrikeSec || 0ldAccounts Suspended @0xSojalSec ||

参加 October 2021

6K フォロー中 55.7K ファン

Md Ismail Šojal 🕷️@0x0SojalSec

2026.08.13 09:06

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, It’s 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. -