Mistral Small 3.1 24B is a 24B-parameter dense model. Every parameter is read for every token generated, so both the memory it occupies and the speed it runs at follow directly from that one figure.
The short answer: at Q4_K_M, the quantisation most people actually run, Mistral Small 3.1 24B occupies roughly 13.5 GB of weights and needs about 16.9 GB of total VRAM once you add an 8K context window and runtime overhead. That puts it within reach of a RX 7900 XTX (24 GB) and anything larger.
Quantisation trades precision for memory. Each row is Mistral Small 3.1 24B at 8K context, including the KV cache and ~1 GB of runtime overhead:
| Quant | Weights | Total VRAM | Quality |
|---|---|---|---|
Q8_0 | 25.5 GB | 28.9 GB | Near-lossless. Use when VRAM is free. |
Q6_K | 19.8 GB | 23.2 GB | Very close to FP16; the safe default if it fits. |
Q5_K_M | 17.1 GB | 20.5 GB | Mild loss, noticeably smaller. |
Q4_K_M | 13.5 GB | 16.9 GB | The sweet spot most people run. |
Q3_K_M | 10.5 GB | 13.9 GB | Visible degradation. Fallback only. |
Below Q4 the returns turn sharply negative. Q3_K_M saves 3.0 GB over Q4_K_M but introduces errors you will notice in structured output and long-form reasoning. If Q4 does not fit, a smaller model at Q5 or Q6 almost always beats this one at Q3.
Every current consumer card, against Mistral Small 3.1 24B at Q4_K_M. Throughput is estimated from memory bandwidth against the full weights:
| GPU | VRAM | Bandwidth | Runs it? | Est. speed |
|---|---|---|---|---|
| RTX 5090 | 32 GB | 1,792 GB/s | Yes — Q8_0 | ~56 tok/s |
| RTX 4090 | 24 GB | 1,008 GB/s | Yes — Q6_K | ~41 tok/s |
| RTX 3090 | 24 GB | 936 GB/s | Yes — Q6_K | ~38 tok/s |
| RX 7900 XTX | 24 GB | 960 GB/s | Yes — Q6_K | ~39 tok/s |
| RTX 5080 | 16 GB | 960 GB/s | Yes — Q3_K_M | ~73 tok/s |
| RTX 5070 Ti | 16 GB | 896 GB/s | Yes — Q3_K_M | ~68 tok/s |
| RTX 5060 Ti 16GB | 16 GB | 448 GB/s | Yes — Q3_K_M | ~34 tok/s |
| RTX 4080 SUPER | 16 GB | 736 GB/s | Yes — Q3_K_M | ~56 tok/s |
| RTX 4070 Ti SUPER | 16 GB | 672 GB/s | Yes — Q3_K_M | ~51 tok/s |
| RTX 4060 Ti 16GB | 16 GB | 288 GB/s | Yes — Q3_K_M | ~22 tok/s |
| RX 9070 XT | 16 GB | 645 GB/s | Yes — Q3_K_M | ~49 tok/s |
| RTX 5070 | 12 GB | 672 GB/s | Needs offload | — |
| Arc B580 | 12 GB | 456 GB/s | Needs offload | — |
Notice that VRAM and bandwidth do not move together. The RTX 4060 Ti 16GB holds as much as an RTX 4080 SUPER but reads it at 288 GB/s against 736 GB/s, so it will load this model and then generate at roughly a third the speed. Capacity decides whether it runs; bandwidth decides whether you enjoy using it.
Fastest path, using Ollama:
ollama run mistral-small-3-1-24bFor control over quantisation and context, pull the GGUF directly and serve it with llama.cpp:
huggingface-cli download mistralai/Mistral-Small-3.1-24B-Instruct-2503 \
--include "*Q4_K_M*.gguf" --local-dir ./models
llama-server -m ./models/*Q4_K_M*.gguf \
--n-gpu-layers 999 \
--ctx-size 8192--n-gpu-layers 999 pushes everything onto the GPU; lower it until the model loads if you are short on memory. --ctx-size is worth tuning deliberately — this model supports up to 128K tokens, but the KV cache grows with it, and asking for the full window when you only need 8K can cost you several gigabytes for nothing.
About 16.9 GB at Q4_K_M with an 8K context window: 13.5 GB of weights plus KV cache and roughly 1 GB of runtime overhead. At Q8_0 it needs 28.9 GB, and at Q3_K_M it comes down to 13.9 GB.
The RX 7900 XTX at 24 GB is the least expensive card that holds it at Q4_K_M, at around $999.
Not entirely. It needs 16.9 GB at Q4_K_M, so an 8 GB card has to offload layers to system RAM, which cuts speed substantially. A smaller model at a higher quantisation will give better results on that hardware.
The weights are published openly on Hugging Face as mistralai/Mistral-Small-3.1-24B-Instruct-2503, but the licence is set by the model's publisher and varies — some are Apache 2.0 or MIT, others carry usage restrictions or revenue thresholds. Check the licence file on the model card before deploying it commercially.
Measurably, but far less than people expect down to Q4. Q6_K is close enough to FP16 that differences are hard to detect; Q4_K_M costs a small amount of accuracy on reasoning and structured output; below Q4 the degradation becomes obvious. Running a larger model at Q4 generally beats a smaller one at Q8.
These figures assume the whole model sits in VRAM and the GPU sustains its rated bandwidth. Real throughput drops if any layers are offloaded to system RAM, if your context is long enough that attention starts to dominate, or if the card is thermally throttling. Prompt processing is also compute-bound rather than bandwidth-bound and follows different limits.
Budget 15 GB for the Q4_K_M GGUF, or 28 GB for Q8_0. Download the specific quantisation you want rather than the whole repository — most GGUF repos hold every variant, and cloning all of them wastes a great deal of space.