Only with CPU offload
Mistral NeMo 12B at Q4_K_M needs 8.7 GB but only 7.0 GB is addressable, so about 25% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 13 tokens per second — usable for batch work, painful for chat.
75% on GPU
8K context
Q4_K_M · 6.9 GB
Apache 2.0
Released Jul 2024
Multilingual 12B with a 128K window, built with NVIDIA. A roleplay and fiction favourite that refuses to die.
What hardware do I need for Mistral NeMo 12B? →
Fits instead: Mistral 7B Instruct v0.3 (5.7 GB)
The VRAM budget
weights 6.9 GB
Weights 6.9 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Over budget 1.7 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 22.7 GB | 24.6 GB | — | ~1.9 | Reference | 17.6 GB over |
| Q8_0 | 12.1 GB | 13.9 GB | — | ~4.5 | −0.1% ppl | 6.9 GB over |
| Q6_K | 9.3 GB | 11.2 GB | — | ~6.9 | −0.4% ppl | 4.2 GB over |
| Q5_K_M | 8.1 GB | 9.9 GB | — | ~9.1 | −0.8% ppl | 2.9 GB over |
| Q4_K_M | 6.9 GB | 8.7 GB | — | ~13 | −1.9% ppl | 1.7 GB over |
| Q3_K_M | 5.6 GB | 7.4 GB | 5K | ~25 | −5.4% ppl | 0.4 GB over |
| Q2_K | 4.8 GB | 6.6 GB | 10K | 37 | −15% ppl | Fits |
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/Mistral-Nemo-Instruct-2407:Q4_K_M \
-c 8192 -ngl 30
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 6.9 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.