For sale: Dual RTX 3090 AI Inference Workstation (Threadripper 3970X)

A dedicated, tuned local-LLM inference box — currently serving Qwen3.8-Flash-Next (125B MoE) 24/7 via llama.cpp, benchmarked and documented. Sold as complete, working system with current inference server setup included (llama.cpp build + launch config, can demo before pickup).

Specs

• *CPU

• Detail: AMD Ryzen Threadripper 3970X — 32C/64T, boost 4.55 GHz


Motherboard

• Detail: ASRock TRX40 Creator (quad-channel DDR4)

*RAM
• Detail: 64 GB DDR4 (4×16 GB)

*GPU
• Detail: 2× NVIDIA RTX 3090 24 GB = 48 GB VRAM (350 W power limit each)

*Storage
• Detail: 2 TB NVMe SSD (Crucial P2 CT2000P2SSD8)

*RAM
• Detail: Corsair Vengeance 4 x 16GB DDR4 3600

*Power Supply
• Detail: Antec 1300W platinum

Why this rig

- 48 GB VRAM in one machine — runs 70B-class dense models at Q6_K and 125B-class MoE models (Qwen3.8-Flash-Next) with expert offloading.
- Threadripper 32 cores handles MoE CPU-offload superbly (that's why Flash-Next hits ~46 tok/s here)
- Currently runs a production llama.cpp server (:18020, OpenAI-compatible API) with 262k context — measured 46.6 tok/s generation / 889 tok/s prefill on Flash-Next IQ3_S
- Fresh OS install (Oct 28), stable uptime, headless-ready

Ideal for: local AI enthusiasts, ML inference homelab, or anyone wanting a turnkey LLM box.