Why Command A pressures system RAM
Command A is a dense 111B network — every weight participates each token, so quantization choice dominates. Q4_K_M lands near ~62.4GB weights, plus ~0.27GB KV at 8K and ~8GB overhead (~70.7GB → 96GB kit). The 256K-token context ceiling is the sleeper cost: long-doc or agent traces inflate KV while the 111B slab stays fixed. Prefer dual-channel DDR5 bandwidth when CPU offload or mmap is involved.
What RAM kit to buy
Buy a matched dual-channel DDR5 kit at 96GB for Command A (EXPO/XMP only if stable). Avoid single-stick installs — local inference is bandwidth-sensitive when layers spill to host memory. Pair with 4x RTX 3090 / 4090 (96GB VRAM) or Apple Mac Studio (128GB Unified Memory) when staying in the Multi-GPU Workstation / Mac Studio tier, and keep 20–30% RAM free for the OS + browser.
Workload notes
Cohere models like Command A are often enterprise-RAG oriented — size RAM for embedding caches and concurrent retrieval workers, not only the LLM weights. At 111B, Command A sits in the large local-LLM band: Q4 on a strong GPU is realistic, FP16 usually is not on consumer cards. Release window noted as 2025/2026; always re-check the model card before buying hardware for a specific checkpoint.