Only with CPU offload

Ministral 3 14B at Q4_K_M needs 9.7 GB but only 8.8 GB is addressable, so about 11% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 26 tokens per second — usable for batch work, painful for chat.

89% on GPU 8K context Q4_K_M · 7.8 GB Apache 2.0 Released Dec 2025 Vision

The largest Ministral. A 12 GB card runs it at Q4 with a few gigabytes to spare.

What hardware do I need for Ministral 3 14B? →

Fits instead: Ministral 3 8B (6.7 GB) · Ministral 3 3B (3.6 GB)

The VRAM budget

weights 7.8 GB
Weights 7.8 GB KV cache @ 8K 1.25 GB Runtime overhead 0.6 GB Over budget 0.9 GB past 8.8 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
F16 25.9 GB 27.7 GB ~1.9 Reference 18.9 GB over
Q8_0 13.8 GB 15.6 GB ~4.9 −0.1% ppl 6.8 GB over
Q6_K 10.6 GB 12.5 GB ~8.6 −0.4% ppl 3.7 GB over
Q5_K_M 9.2 GB 11.0 GB ~13 −0.8% ppl 2.2 GB over
Q4_K_M 7.8 GB 9.7 GB 2K ~26 −1.9% ppl 0.9 GB over
Q3_K_M 6.3 GB 8.2 GB 11K 73 −5.4% ppl Fits
Q2_K 5.4 GB 7.3 GB 17K 85 −15% ppl Long context

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.

How to run it

terminal
$ llama-server \
    -hf mistralai/Ministral-3-14B-Instruct-2512:Q4_K_M \
    -c 8192 -ngl 35

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 7.8 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.
See all models for this rig Compare with another model