VRAM Fit Checker: Will It Run Your Model?

Pick an open model and a precision. The checker estimates the memory the weights need and lists every GPU, mini PC, workstation and server in our directory that can hold it.

Model Precision

How the estimate works: parameters x bytes per parameter x 1.2, where the extra 20 percent covers the KV cache and runtime overhead for a typical chat workload. Long contexts need more. Two free checks a day; after that, an email unlocks the checker for the rest of the day.

Your results land here: the memory needed, and the hardware in the directory that fits it, tightest fit first.

The reference table behind the checker

The same math, printed out for the models in the checker.

Model	Params	FP16 needs	INT8 needs	INT4 needs	Note
Llama 3.1 8B	8B	19.2 GB	9.6 GB	4.8 GB	Dense model
Llama 3.1 70B	70B	168 GB	84 GB	42 GB	Dense model
Llama 3.1 405B	405B	972 GB	486 GB	243 GB	Dense model
Qwen 2.5 7B	7B	16.8 GB	8.4 GB	4.2 GB	Dense model
Qwen 2.5 32B	32B	76.8 GB	38.4 GB	19.2 GB	Dense model
Qwen 2.5 72B	72B	172.8 GB	86.4 GB	43.2 GB	Dense model
Mistral 7B	7B	16.8 GB	8.4 GB	4.2 GB	Dense model
Mixtral 8x7B	46.7B	112.08 GB	56.04 GB	28.02 GB	MoE: all experts must fit in memory
DeepSeek V3	671B	1.61 TB	805.2 GB	402.6 GB	MoE: 37B active per token, but all 671B of weights must fit
DeepSeek R1	671B	1.61 TB	805.2 GB	402.6 GB	MoE: 37B active per token, but all 671B of weights must fit
FLUX.1 dev	12B	28.8 GB	14.4 GB	7.2 GB	Image model: estimate covers the transformer, add room for the text encoders
SDXL	3.5B	8.4 GB	4.2 GB	2.1 GB	Image model: small weights, activations grow with resolution