Can I Run Mistral Family on 64 GB system RAM?

Superseded model. Mistral Family has been superseded by Mistral Small 4. This page is kept for reference; the newer family is a better starting point. View Mistral Small 4 →

Written by Jakub Rusinowski · Last updated January 28, 2025

Yes — comfortably

Yes, comfortably — Mistral Small 3 (24B) at Q3_K_M needs about 12.9 GB of the 51.2 GB usable on 64 GB system RAM, leaving ~38.3 GB spare and running at ~5.9 tok/s (estimated), with room for about 32,768 tokens of context.

Confidence: high · Recommended quantization: Q3_K_M · Estimated speed: ~5.9 tok/s

64 GB system RAM — what it gives a model

Usable memory for models	51.2 GB
Memory bandwidth	90 GB/s

Mistral Family on 64 GB system RAM: memory by quantization

Quant	Memory needed	Fits 51.2 GB?	Max context	Est. speed	Download
F16	50.7 GB	✓ Yes	8K	~1.4 tok/s	48 GB
Q8_0	28.2 GB	✓ Yes	32K	~2.5 tok/s	25.5 GB
Q6_K	22.4 GB	✓ Yes	32K	~3.2 tok/s	19.7 GB
Q5_K_M	19.7 GB	✓ Yes	32K	~3.7 tok/s	17 GB
Q4_K_M	17.2 GB	✓ Yes	32K	~4.3 tok/s	14.5 GB
Q3_K_M	12.9 GB	✓ Yes	32K	~5.9 tok/s	10.2 GB
Q2_K	10.6 GB	✓ Yes	32K	~7.5 tok/s	7.9 GB

Which Mistral Family sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Mistral Small 3 (24B)	17.2 GB	✓ Fits	~4.3 tok/s
Mistral NeMo 12B	9.4 GB	✓ Fits	~8.3 tok/s

What to watch out for

Q3_K_M 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.
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.
51.2 GB of the 64 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