Can I Run Ministral on 8 GB system RAM?

Superseded model. Ministral has been superseded by Mistral Small 4.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 October 16, 2024

Yes

Yes — Ministral 3B at Q8_0 needs about 5.5 GB of the 6.4 GB usable on 8 GB system RAM (~0.9 GB spare), at ~15.2 tok/s (estimated), with room for about 16,384 tokens of context.

Confidence: medium · Recommended quantization: Q8_0 · Estimated speed: ~15.2 tok/s

8 GB system RAM — what it gives a model

Usable memory for models	6.4 GB
Memory bandwidth	90 GB/s

Ministral on 8 GB system RAM: memory by quantization

Quant	Memory needed	Fits 6.4 GB?	Max context	Est. speed	Download
F16	8.9 GB	✗ No	—	—	7.2 GB
Q8_0	5.5 GB	✓ Yes	16K	~15.2 tok/s	3.8 GB
Q6_K	4.6 GB	✓ Yes	16K	~19 tok/s	3 GB
Q5_K_M	4.2 GB	✓ Yes	16K	~21.4 tok/s	2.6 GB
Q4_K_M	3.8 GB	✓ Yes	16K	~24.3 tok/s	2.2 GB
Q3_K_M	3.2 GB	✓ Yes	32K	~31.7 tok/s	1.5 GB
Q2_K	2.9 GB	✓ Yes	32K	~38 tok/s	1.2 GB

Which Ministral sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Ministral 8B	6.9 GB	✗ Too large	—
Ministral 3B	3.8 GB	✓ Fits	~24.3 tok/s

What to watch out for

Only ~0.9 GB of headroom at Q8_0: a longer context or a second application can push this into swapping.
1 larger variant of Ministral does not fit and would need CPU offload or different hardware.
This model does not publish its full attention configuration, so the KV-cache share of these figures is inferred from its parameter count rather than computed exactly.
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.
6.4 GB of the 8 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 inferred from the parameter count because this model does not publish its full attention configuration.

FAQ