Why Saba pressures system RAM
Saba is a dense 24B network — every weight participates each token, so quantization choice dominates. Q4_K_M lands near ~13.5GB weights, plus ~0.06GB KV at 8K and ~6GB overhead (~19.6GB → 32GB kit). The 33K-token context ceiling is the sleeper cost: long-doc or agent traces inflate KV while the 24B slab stays fixed. Prefer dual-channel DDR5 bandwidth when CPU offload or mmap is involved.
What RAM kit to buy
A 32GB dual-channel kit is enough for quantized Saba at modest context. Still prefer 2× matched SO-DIMM/UDIMM sticks; 1x RTX 3090 or RTX 4090 (24GB VRAM) covers the Flagship Consumer GPU GPU profile. If you chat with long pastes, jump a tier before the KV cache forces paging.
Workload notes
Mistral releases like Saba are common in production vLLM; dual-channel bandwidth helps prompt throughput when CPU offload is in play. At 24B, Saba 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.