← Interactive LLM VRAM Calculator — try your own context length & quantization

How much VRAM to run Mixtral 8x7B? (FP16/Q8/Q6/Q4 requirements)

Architecture: Mixture-of-Experts: 46.7B total / ~12.9B active parameters, 32 layers, 8 KV heads. Figures are approximations for weights + an 8k-token FP16 KV cache + ~1 GB runtime overhead, based on typical published GGUF sizes. Actual usage varies by runtime and settings.

QuantizationWeightsTotal VRAM (8k context)Fits on (example)
FP16
full precision (16-bit)
~93 GB~95 GB2×RTX A6000 (96 GB)
Q8_0
8-bit
~50 GB~52 GB2×RTX 5090 (64 GB)
Q6_K
~6.6-bit
~38 GB~40 GB2×RTX 3090 or RTX A6000 (48 GB)
Q5_K_M
~5.7-bit
~33 GB~35 GB2×RTX 3090 or RTX A6000 (48 GB)
Q4_K_M
~4.8-bit
~26 GB~28 GBRTX 5090 (32 GB)

How much VRAM do I need to run Mixtral 8x7B?

You need roughly ~28 GB of VRAM to run Mixtral 8x7B at Q4_K_M with an 8k-token context — that fits on RTX 5090 (32 GB). At Q8_0 plan for about ~52 GB, and full FP16 needs around ~95 GB. Mixtral is a Mixture-of-Experts model: only ~12.9B parameters are active per token (so it is fast), but all 46.7B must sit in memory. Fully on-GPU Q4_K_M needs ~29 GB; many people run it on a 24 GB card by offloading some layers to CPU RAM with llama.cpp, at reduced speed.

What affects these numbers?

Three things: the quantization (bytes per parameter), the KV cache (grows linearly with context length — the table assumes 8k tokens), and runtime overhead. Longer contexts need more memory; quantizing the KV cache (e.g. q8_0 cache in llama.cpp) roughly halves that part. Use the interactive calculator to plug in your exact context length and quantization.