Why DeepSeek-R1-0528 (671B MoE) pressures system RAM
DeepSeek-R1-0528 (671B MoE) is Mixture-of-Experts: inference activates 37B active/token, but VRAM/RAM must usually hold the full ~671B expert set for fast routing. At Q4 the weight slab is ~377.4GB before KV (~0.09GB at 8K) and ~12GB OS/runtime overhead — totaling ~389.5GB raw, rounded to a 512GB kit. Stretching toward the full 128K-token window multiplies KV far faster than weights; that is the usual “I bought enough RAM for the model but still OOM” failure on DeepSeek MoE pages.
What RAM kit to buy
Shop 512GB-class capacity for DeepSeek-R1-0528 (671B MoE): workstation DDR5 RDIMM/LRDIMM or multi-kit desktop builds, not a single gamer 2×16GB stick. Use our 128GB+ price hubs and RAM Finder; confirm ECC needs for your board. GPU path: Apple Mac Studio (192GB Unified Memory) or Institutional Node (8x H100 / A100) (389.4GB VRAM class) if you want weights on-device instead of system-RAM offload.
Workload notes
DeepSeek checkpoints such as DeepSeek-R1-0528 (671B MoE) are popular in GGUF community quants; watch for sparse-attention / MLA variants that change KV growth vs plain dense transformers. At 671B total parameters this is frontier-scale — expect multi-GPU or heavy CPU offload even in Q4; the 512GB kit is a host-memory floor, not a promise of interactive tokens/s. Release window noted as May 2025; always re-check the official source before buying hardware for a specific checkpoint.