RunLLMLocal
Open the checker

Models · chat and writing, image understanding

Mistral Small 3 24B: VRAM requirements and GPUs that run it

Fast for its size, vision capable, Apache licence. At Q4 with 8K context it needs about 18 GB, so you want a GPU with at least 20 GB of VRAM or a Mac with 24 GB of unified memory.

Can your PC run Mistral Small 3 24B?

Pick your graphics card or Mac. Mistral Small 3 24B is pinned at the top of the list.

Models 0 of 0 run

● Runs in your browserMemory = weights × 1.1 + KV cache. Usable = 90% VRAM, 72% on Mac.
24Bparameters
18 GBat Q4, 8K context
5/40GPUs run it fully
~59tok/s, fastest card

Memory needed by quantization

QuantWeightsTotal at 8KTotal at 32K
Q4_K_M16 GB18 GB22 GB
Q5_K_M19 GB21 GB26 GB
Q8_028 GB30 GB35 GB
FP1653 GB55 GB60 GB

Totals include about 10% runtime overhead and the KV cache, plus around 0.6 GB for the vision encoder. Longer conversations cost real memory: going from 8K to 32K context adds about 4.7 GB at any quantization. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.

Which GPUs can run Mistral Small 3 24B?

5 of the 40 cards we track run it fully in VRAM at 8K context. The cheapest entry point is a 20 GB card such as the RX 7900 XT, and the fastest is the RTX 5090 at about 59 tokens per second. 35 more can run it with part of the model in system RAM, which is noticeably slower.

GPUVRAMStatusQuantSpeed
RTX 509032 GBRuns wellQ6_K~59 tok/s
RTX 409024 GBRuns wellQ5_K_M~38 tok/s
RX 7900 XT20 GBRuns wellQ4_K_M~37 tok/s
RX 7900 XTX24 GBRuns wellQ5_K_M~37 tok/s
RTX 3090 / 3090 Ti24 GBRuns wellQ5_K_M~36 tok/s
RTX 508016 GBOffloadQ4_K_M~5.8 tok/s
RTX 5070 Ti16 GBOffloadQ4_K_M~5.5 tok/s
RTX 5060 Ti16 GBOffloadQ4_K_M~3.0 tok/s
RTX 4080 / 4080 Super16 GBOffloadQ4_K_M~4.6 tok/s
RTX 4070 Ti Super16 GBOffloadQ4_K_M~4.2 tok/s
RTX 4060 Ti16 GBOffloadQ4_K_M~2.1 tok/s
RTX 4090 Laptop16 GBOffloadQ4_K_M~3.7 tok/s
RX 9070 XT16 GBOffloadQ4_K_M~4.0 tok/s
RX 7800 XT16 GBOffloadQ4_K_M~4.0 tok/s
RX 6800 XT16 GBOffloadQ4_K_M~3.3 tok/s
Arc A77016 GBOffloadQ4_K_M~3.6 tok/s
RTX 507012 GBOffloadQ4_K_M~3.8 tok/s
RTX 4070 / 4070 Super12 GBOffloadQ4_K_M~3.1 tok/s
RTX 306012 GBOffloadQ4_K_M~2.5 tok/s
RTX 4080 Laptop12 GBOffloadQ4_K_M~2.8 tok/s
RX 7700 XT12 GBOffloadQ4_K_M~2.8 tok/s
RX 6700 XT12 GBOffloadQ4_K_M~2.6 tok/s
Arc B58012 GBOffloadQ4_K_M~2.9 tok/s
GTX 1080 Ti11 GBOffloadQ4_K_M~3.0 tok/s
RTX 308010 GBOffloadQ4_K_M~3.9 tok/s
RTX 5060 Ti8 GBOffloadQ4_K_M~2.7 tok/s
RTX 50608 GBOffloadQ4_K_M~2.7 tok/s
RTX 4060 Ti8 GBOffloadQ4_K_M~2.3 tok/s
RTX 40608 GBOffloadQ4_K_M~2.2 tok/s
RTX 3070 / 3070 Ti8 GBOffloadQ4_K_M~2.7 tok/s
RTX 3060 Ti8 GBOffloadQ4_K_M~2.7 tok/s
RTX 30508 GBOffloadQ4_K_M~2.1 tok/s
RTX 4070 Laptop8 GBOffloadQ4_K_M~2.2 tok/s
RTX 4060 Laptop8 GBOffloadQ4_K_M~2.2 tok/s
RX 76008 GBOffloadQ4_K_M~2.3 tok/s
Arc A7508 GBOffloadQ4_K_M~2.9 tok/s
GTX 1660 / 1660 Super6 GBOffloadQ4_K_M~2.4 tok/s
RTX 4050 Laptop6 GBOffloadQ4_K_M~2.1 tok/s
RTX 3060 Laptop6 GBOffloadQ4_K_M~2.4 tok/s
GTX 16504 GBOffloadQ4_K_M~2.1 tok/s

Running it on a Mac

Apple Silicon Macs can give a model about 72% of their unified memory. Mistral Small 3 24B fits from 24 GB (Apple M3, about 5.6 tokens per second). With a faster chip like the Apple M3 Ultra (96 GB) expect around 21 tokens per second. Compare the options in our Mac vs NVIDIA guide.

Similar size models: Phi-4 14B, Qwen3 14B, DeepSeek-R1 Distill 14B, gpt-oss 20B, Codestral 22B.

Compare Mistral Small 3 24B

Mistral Small 3 24B vs Gemma 3 27BSide by side for local LLMsCompare with any modelPick the other one yourself

Questions

How much VRAM does Mistral Small 3 24B need?

About 18 GB at Q4_K_M with 8K context, 30 GB at Q8_0, and 55 GB at full FP16. In practice that means a card with at least 20 GB of VRAM for Q4.

Can Mistral Small 3 24B run on an 8 GB GPU?

Only partly. It needs about 18 GB at Q4, so the rest spills into system RAM and speed drops sharply.

Can I run Mistral Small 3 24B on a Mac?

Yes. The smallest configuration that runs it fully is Apple M3 with 24 GB of unified memory, at about 5.6 tokens per second. More memory bandwidth (Pro, Max, Ultra chips) means faster replies.

Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.

Copied