$3000 home server with 128GB VRAM runs Qwen3.8-next at 1.3k tps prefill, 70 tps code
Thin_Pollution8843 · reddit · 2026-09-14
- A Redditor finished a home inference server costing $3000 (SSD excluded): 128GB VRAM + 256GB DDR4 RAM.
- Build: 4x RTX V620 ($1400), 256GB DDR4 RDIMM 2666 ($610), Huanandzhi D12D board ($410), EPYC 7452 ($170), ASRock 1600W PSU ($220), case/fans/misc $200. An earlier Lenovo P620 attempt was returned over proprietary parts.
- Power draw is real: 700–900W prefill, 500–600W decode running Qwen3.8-next-flash Autoround W4A16.
- Results: with a vllm fork + MTP-2 at 128k+ context, 1.3k tps prefill and 70 tps code / 60 tps prose generation. The author was initially disappointed with Qwen3.8-27b speeds but is satisfied with Qwen3.8-next, hoping for more optimizations.
More from Infra
- Would LITE accepting MRVL's merger offer prove its moat was overstated? — BenBajarin · 2026-10-30
- Claymation body horror on 8GB: three wrong settings, not the GPU, made local renders slow — roro9- · 2026-10-30
- durable-object-migrator: zero-downtime exports for Cloudflare Durable Objects — samgoodwin89 · 2026-10-30
- An unnamed firm already runs a fairly scaled dedicated vehicle for compute lending, says analyst — matt_slotnick · 2026-10-30
- GPU break-even throughput vs open-weight token prices up 25-83% since May — bookwormengr · 2026-10-30
- Sam Altman says the chip supply chain's invisible hand is failing — and DRAM is why — SnoozeDoggyDog · 2026-10-30