Can I Run Command R Family on 32 GB system RAM?

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 →

Written by Jakub Rusinowski · Last updated April 4, 2024

Technically yes, but not recommended

It loads, but it is not worth running — Command R (35B) at Q2_K fits in 32 GB system RAM's 25.6 GB, yet the memory bandwidth limits it to ~3.9 tok/s (estimated), well below usable interactive speed.

Confidence: high · Recommended quantization: Q2_K · Estimated speed: ~3.9 tok/s

32 GB system RAM — what it gives a model

Usable memory for models	25.6 GB
Memory bandwidth	90 GB/s

Command R Family on 32 GB system RAM: memory by quantization

Quant	Memory needed	Fits 25.6 GB?	Max context	Est. speed	Download
F16	81.5 GB	✗ No	—	—	70 GB
Q8_0	48.7 GB	✗ No	—	—	37.2 GB
Q6_K	40.2 GB	✗ No	—	—	28.7 GB
Q5_K_M	36.3 GB	✗ No	—	—	24.8 GB
Q4_K_M	32.7 GB	✗ No	—	—	21.1 GB
Q3_K_M	26.5 GB	✗ No	—	—	14.9 GB
Q2_K	23 GB	✓ Yes	8K	~3.9 tok/s	11.5 GB

Which Command R Family sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Command R+ (104B)	65.7 GB	✗ Too large	—
Command R (35B)	32.7 GB	✗ Too large	—

What to watch out for

At ~3.9 tok/s this loads but is too slow for interactive use — expect roughly 15 seconds per 60 tokens.
Q2_K is the only quantization that fits, and it is a heavily degraded one — expect noticeably worse output than the same model at Q4_K_M. A smaller model at Q4 is usually the better trade.
2 larger variants of Command R Family do not fit and would need CPU offload or different hardware.
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.
25.6 GB of the 32 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