Can I Run Command R Family on 96 GB system RAM?

Autor: Jakub Rusinowski · Ostatnia aktualizacja: 4 kwietnia 2024

Technically yes, but not recommended

It loads, but it is not worth running — Command R+ (104B) at Q5_K_M fits in 96 GB system RAM's 76.8 GB, yet the memory bandwidth limits it to ~0.9 tok/s (estimated), well below usable interactive speed.

Confidence: high · Recommended quantization: Q5_K_M · Estimated speed: ~0.9 tok/s

Get a personalized upgrade path →

Kup ten sprzęt NVIDIA DGX Spark (128GB) — 128 GB VRAM · 150 W board power Uruchom w chmurze NVIDIA A100 80GB w RunPod — od $1.39/godz. · stawka sprawdzona 2026-07

lub porównaj w Vast.ai od $0.77/godz. (typical low · varies)

Jako partner Amazon zarabiamy na kwalifikujących się zakupach. Linki do chmurowych GPU są linkami polecającymi — możemy otrzymać prowizję bez dodatkowych kosztów dla Ciebie.

96 GB system RAM — what it gives a model

Usable memory for models	76.8 GB
Memory bandwidth	90 GB/s

Command R Family on 96 GB system RAM: memory by quantization

Quant	Memory needed	Fits 76.8 GB?	Max context	Est. speed	Download
F16	210.9 GB	✗ No	—	—	208 GB
Q8_0	113.4 GB	✗ No	—	—	110.5 GB
Q6_K	88.2 GB	✗ No	—	—	85.3 GB
Q5_K_M	76.7 GB	✓ Yes	8K	~0.9 tok/s	73.7 GB
Q4_K_M	65.7 GB	✓ Yes	32K	~1 tok/s	62.8 GB
Q3_K_M	47.3 GB	✓ Yes	64K	~1.5 tok/s	44.3 GB
Q2_K	37.1 GB	✓ Yes	64K	~1.9 tok/s	34.2 GB

Newer alternative. Cohere has not released a newer model in this library. If you are choosing today, Mistral Small 4 is a comparable current option. View Mistral Small 4 119B-A6.5B →

Which Command R Family sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Command R+ (104B)	65.7 GB	✓ Fits	~1 tok/s
Command R (35B)	32.7 GB	✓ Fits	~2.5 tok/s

What to watch out for

At ~0.9 tok/s this loads but is too slow for interactive use — expect roughly 67 seconds per 60 tokens.
Only ~0.1 GB of headroom at Q5_K_M: a longer context or a second application can push this into swapping.
These figures assume CPU-only inference. Any discrete GPU, even an 8 GB one, will be several times faster for models that fit in its VRAM.

Recommended setup

llama.cpp (CPU build) or Ollama — both run without a GPU

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
76.8 GB of the 96 GB is treated as usable for model weights (80% — the rest is the OS and running applications).
DDR5-5600 dual channel at 89.6 GB/s peak. CPU decode is assumed to sustain 35% of that peak, because CPU inference is not purely bandwidth-bound — it also spends real time in compute and thread synchronisation. This figure is an assumption, not a fitted constant: no CPU measurement is in the calibration set.
CPU-only inference: no GPU is assumed. A GPU of any size will beat these figures substantially.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is computed from this model's published attention configuration.

FAQ