Yes, just — 0.9 GB spare
Ministral 3 14B at Q4_K_M fits your GeForce RTX 2060 12 GB entirely on the GPU at 8K context, at an estimated 26 tokens per second. Past 13K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully 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.
The VRAM budget
weights 7.8 GB
Weights 7.8 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Free 0.9 GB of 10.6 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 25.9 GB | 27.7 GB | — | ~1.9 | Reference | 17.1 GB over |
| Q8_0 | 13.8 GB | 15.6 GB | — | ~5.5 | −0.1% ppl | 5.0 GB over |
| Q6_K | 10.6 GB | 12.5 GB | — | ~11 | −0.4% ppl | 1.9 GB over |
| Q5_K_M | 9.2 GB | 11.0 GB | 5K | ~18 | −0.8% ppl | 0.4 GB over |
| Q4_K_M | 7.8 GB | 9.7 GB | 13K | 26 | −1.9% ppl | Recommended |
| Q3_K_M | 6.3 GB | 8.2 GB | 23K | 32 | −5.4% ppl | Long context |
| Q2_K | 5.4 GB | 7.3 GB | 29K | 38 | −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
$ llama-server \
-hf mistralai/Ministral-3-14B-Instruct-2512:Q4_K_M \
-c 8192 -ngl 99
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.
03Only 0.9 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 13K context.