48GB RTX 4090 Local-LLM Server

The listing

High-end local-inference rig, sold barebones minus storage (no SSD, no OS — add any NVMe and install Linux; ~30 min and you're running). ★ NVIDIA RTX 4090 with 48GB VRAM (aftermarket VRAM mod). Verified on this exact card: Llama-3.3-70B (q4): 21.9 tok/s - 100% in VRAM (43GB / 48GB), no CPU offload Qwen 27B (q8): 27.7 tok/s 256GB system RAM allows offload to ~120B quantized Specs: GPU: RTX 4090, 48GB VRAM (modded) CPU: Intel i9-10980XE - 18C / 36T RAM: 256GB DDR4-3200MHZ NON ECC 3200MHZ (8×32GB) Board: ASUS PRIME X299-A II Storage: NONE included - bring your own NVMe SSD OS: not included - installs clean (validated on Ubuntu 24.04 + Ollama; use any modern Linux) Condition - full transparency: 48GB VRAM-modded 4090 (aftermarket; no mfg warranty on the mod). Validated stable at STOCK 450W - 10-min continuous 70B soak held 449W @ 78°C, full clocks, zero crash. Fresh thermal paste + VRAM pads.