2026-10-17 · America/Los_Angeles · 社区动态 · #4

You don’t need that expensive of a build for local llms.

Been seeing a lot of people with their 64gb of ddr5 4x3090 or whatever strata builds and it’s been sorta depressing knowing i cant afford that but i finally finished my setup today so here we go, Qwen 3.8 27b at 262k context, unsloth iq4XS ascii condensed with kv cache streaming My only gpu is a 9060 xt 16gb and i have 24gb ddr5 for my whole homelab only needing 16 for this MTP+ ngram prediction q4 kv cache values with q8 keys 42.7 tps generation 658 tps prefill at 20k, 35/500tps at 200k Using this fork: https://github.com/achyudh/llama.cpp I got grok bot to ssh in and fix the bugs i had…

热度 47.4 / 100；排名与评分保留该期记录。