Adding €300 of used RAM for 80GB to run MoE models: is it worth it?
quadra-lab · reddit · 2026-10-14
Running Qwen 27B at IQ4 on a 7900XT (20GB VRAM, 40 tps with MTP), the author wants higher-accuracy Q8 MoE models (like Qwen Flash Next) that use system RAM plus GPU. They found used 2×32GB DDR4 3200 CL16 for €300, bringing total RAM to 80GB, and ask whether MoEs really beat lower-quant dense models and for real gaming-rig-to-AI-server setups.
More from Infra
- Hugging Face to livestream how speculative decoding delivers multipliers in local LLM speed-ups — mervenoyann · 2026-10-14
- The Jensen Huang Playbook: How NVIDIA Built CoreWeave as a Buffer Against Its Biggest Customers — AccBalanced · 2026-10-14
- 64GB Nvidia Spark spotted at €6,574 in Europe — €900 more buys the 128GB model — AIFlow_ML · 2026-10-14
- Toby Ord says RL post-training compute neared parity with pretraining; bet on broad compute-capability trends — ben_j_todd · 2026-10-14
- Fractile CEO: smarter AI can guide chip design, but experiments are now the bottleneck — No Priors · 2026-10-14
- China routes surplus solar and wind to AI compute via "East Data, West Computing" — ingliguori · 2026-10-14