Local LLM rig under $4k (2026)

The minimum-viable workstation for serious local inference: single 5090, 64 GB system RAM, fast NVMe, and a case that doesn't give up by hour two.

const{Fragment:e,jsx:n,jsxs:r}=arguments[0];function _createMdxContent(t){const i={h2:"h2",li:"li",p:"p",strong:"strong",table:"table",tbody:"tbody",td:"td",th:"th",thead:"thead",tr:"tr",ul:"ul",...t.components};return r(e,{children:[n(i.h2,{children:"The job"}),"\n",n(i.p,{children:"You want to run 70B-class models at home for development, research, or\nwriting — at usable speed, with a context window that doesn't force you to\nchunk everything. You're allergic to the monthly cloud bill. You have ~$4k\nto spend and you want the rig to still feel fast in eighteen months."}),"\n",n(i.p,{children:"This guide is not for:"}),"\n",r(i.ul,{children:["\n",n(i.li,{children:"Fine-tuning from scratch (you need more VRAM or a multi-GPU rig)."}),"\n",n(i.li,{children:"Pure image/video generation (different tradeoffs, covered in a separate\nguide)."}),"\n",n(i.li,{children:"Production inference serving (this is a workstation, not a datacenter node)."}),"\n"]}),"\n",n(i.h2,{children:"The build"}),"\n",r(i.table,{children:[n(i.thead,{children:r(i.tr,{children:[n(i.th,{children:"Part"}),n(i.th,{children:"Pick"}),n(i.th,{children:"Why"})]})}),r(i.tbody,{children:[r(i.tr,{children:[n(i.td,{children:"GPU"}),n(i.td,{children:"NVIDIA RTX 5090 (32 GB)"}),n(i.td,{children:"70B Q4 with headroom; long context works."})]}),r(i.tr,{children:[n(i.td,{children:"CPU"}),n(i.td,{children:"AMD Ryzen 9 9950X or similar 16-core"}),n(i.td,{children:"You'll bottleneck on single-thread + some lanes."})]}),r(i.tr,{children:[n(i.td,{children:"RAM"}),n(i.td,{children:"64 GB DDR5-6000 (2×32)"}),n(i.td,{children:"Leaves room for KV-cache spill + tooling."})]}),r(i.tr,{children:[n(i.td,{children:"Storage"}),n(i.td,{children:"2 TB PCIe 4.0 NVMe"}),n(i.td,{children:"Model weights + datasets + Docker images."})]}),r(i.tr,{children:[n(i.td,{children:"PSU"}),n(i.td,{children:"1000 W 80+ Gold, single rail"}),n(i.td,{children:"5090 is serious; don't be clever here."})]}),r(i.tr,{children:[n(i.td,{children:"Case"}),n(i.td,{children:"Airflow-first mid-tower; 3× intake / 2× exh."}),n(i.td,{children:"Sustained loads run for hours."})]}),r(i.tr,{children:[n(i.td,{children:"OS"}),n(i.td,{children:"Windows 11 Pro or Ubuntu 24.04"}),n(i.td,{children:"Your call. Both work; drivers are mature."})]})]})]}),"\n",n(i.h2,{children:"Numbers"}),"\n",n(i.p,{children:"Approximate inference throughput on this build with llama.cpp, 70B Q4 quant,\nshort prompt:"}),"\n",r(i.ul,{children:["\n",r(i.li,{children:[n(i.strong,{children:"~18–24 tok/s"})," streaming generation."]}),"\n",r(i.li,{children:[n(i.strong,{children:"~32k context"})," usable without spill."]}),"\n",r(i.li,{children:[n(i.strong,{children:"Cold start"})," dominated by model load from NVMe (~4 seconds for a 70B Q4)."]}),"\n"]}),"\n",n(i.p,{children:"Your mileage will vary with prompt shape and sampler choice."}),"\n",n(i.h2,{children:"Tradeoffs"}),"\n",r(i.ul,{children:["\n",r(i.li,{children:[n(i.strong,{children:"Dual 4090 instead of a single 5090."})," Higher aggregate VRAM (48 GB), but\nyou lose the clean single-card setup, and a lot of local-inference tooling\ndoesn't cleanly split across two cards without effort."]}),"\n",r(i.li,{children:[n(i.strong,{children:"Threadripper instead of Ryzen 9."})," More PCIe lanes, more cores, more money.\nIf you'll add a second GPU in year two, worth it. If not, skip."]}),"\n",r(i.li,{children:[n(i.strong,{children:"Cloud on-demand."})," Breaks even with this rig around ~18 months of heavy\nuse, depending on your cloud tier."]}),"\n"]}),"\n",n(i.h2,{children:"What this doesn't get you"}),"\n",r(i.ul,{children:["\n",n(i.li,{children:"Multi-GPU training. You need NVLink, more lanes, more PSU headroom."}),"\n",n(i.li,{children:"Proper datacenter-style serving (batching, multi-user concurrency)."}),"\n",n(i.li,{children:"A good excuse. Buy the rig."}),"\n"]})]})}return{default:function(e={}){const{wrapper:r}=e.components||{};return r?n(r,{...e,children:n(_createMdxContent,{...e})}):_createMdxContent(e)}};