Can the NVIDIA RTX 3090 run Mixtral 8x7B?

Yes, with heavy quantization
QuantQ2_KHeavily degraded
Memory19.2 GBof 24 GB
Speed~116tokens/s

Only just. Mixtral 8x7B fits on the NVIDIA RTX 3090 at Q2_K, which costs noticeable quality. Expect about 116 tokens/second. A card with more memory is a better experience.

Try other context lengths in the calculator →
Run it in the cloud

Want full quality? Run Mixtral 8x7B on a bigger rented GPU.

Rent a cloud GPU by the hour. No upfront cost, cancel any time.

RunPodOn-demand GPUs with ready-made templates for Ollama and vLLM.Rent a GPU on RunPod → Vast.aiA marketplace of rentable GPUs, from RTX cards to H100s.Rent a GPU on Vast.ai →

Memory needed by quantization

At 8,192 tokens of context.

QuantQualityMemoryFits?Speed
FP16Full quality95.1 GBToo big—
Q8_0Near-lossless51.3 GBToo big—
Q6_KExcellent40 GBToo big—
Q5_K_MVery good34.9 GBToo big—
Q4_K_MGood, most popular30 GBToo big—
Q3_K_MNoticeably degraded24.5 GBToo big—
Q2_KHeavily degraded19.2 GBFits~116 tok/s

How to run Mixtral 8x7B on the NVIDIA RTX 3090

Ollama's default mixtral:8x7b download is Q4_K_M, which needs about 30 GB, more than the 24 GB available. Download the Q2_K GGUF instead (LM Studio lists quants per model), or try a smaller context length.

GPUs that run Mixtral 8x7B comfortably

Check NVIDIA RTX 5090 prices →

Models that run well on the NVIDIA RTX 3090