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,
Proof1 community thread
Key points
- 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
- 42.7 tps generation 658 tps prefill at 20k, 35/500tps at 200k
- I got grok bot to ssh in and fix the bugs i had with missing pcie atomics
Sources (1)
- [1]You don’t need that expensive of a build for local llms.r/LocalLLaMA (top, daily) · Oct 18, 02:46 AM
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
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Basalt: NVFP4 + faster prefill for Qwen3.8 Flash-Next on Blackwell (single 5090)
- Oct 17, 2026Qwen 27B (IQ4_XS, Q8/Q4 KV) at full 262k context on a 16GB RX 9070 XT: 80+ t/s decode with Adaptive KV Streaming
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026What a £300, 35W Snapdragon mini PC can actually do 👀: 35B coder at ~30 tok/s, and an 80B model that doesnt even fit in RAM running at 17 tok/s
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
