Why Mistral Large 3 2512 pressures system RAM
Mistral Large 3 2512 is Mixture-of-Experts: inference activates 5.1B active/token, but VRAM/RAM must usually hold the full ~41B expert set for fast routing. At Q4 the weight slab is ~23.1GB before KV (~0.01GB at 8K) and ~6GB OS/runtime overhead — totaling ~29.1GB raw, rounded to a 32GB kit. Stretching toward the full 262K-token window multiplies KV far faster than weights; that is the usual “I bought enough RAM for the model but still OOM” failure on Mistral AI MoE pages.
What RAM kit to buy
A 32GB dual-channel kit is enough for quantized Mistral Large 3 2512 at modest context. Still prefer 2× matched SO-DIMM/UDIMM sticks; 2x RTX 3090 (48GB combined VRAM) or Mac Studio 64GB covers the Dual GPU Setup (48GB VRAM) GPU profile. If you chat with long pastes, jump a tier before the KV cache forces paging.
Workload notes
Mistral releases like Mistral Large 3 2512 are common in production vLLM; dual-channel bandwidth helps prompt throughput when CPU offload is in play. At 41B, Mistral Large 3 2512 is a practical mid-size local model — sweet spot for single-GPU Q4/Q8 experimenters who still want headroom for IDE + Docker. Release window noted as 2025/2026; always re-check the model card before buying hardware for a specific checkpoint.