How much VRAM does MusicGen Large need?

At BF16 with everything resident, MusicGen Large peaks at about 7.55 GB. Offloading components one at a time brings that to 7.55 GB, and quantising the weights to 4-bit brings it to 3.09 GB.

What is the cheapest GPU that runs MusicGen Large?

The GeForce GTX 1650 at around $100, running it at NF4 with everything resident, taking about 2 min 10 s per generation.

MusicGen · Autoregressive · Music and sound

MusicGen Large

MusicGen Large is a single network of 3.3 billion parameters, of which 3.3 billion do the actual generating. It generates one audio token at a time, which makes it bandwidth-bound like a language model rather than compute-bound like a diffusion model. What sets its speed is not its size but its token rate: 50 tokens for every second of output.

Whole pipeline3.3B

Denoiser3.3B

Download12.3 GB

What the pipeline is made of

This is the part that separates a media model from a language model: these are separate networks, and whether they sit in memory together is your choice, not the model's.

Component	Parameters	Published as	On disk	What it does
model	3.3B	F32	12.29 GB	Auxiliary encoder.

These figures were stated by us, because the weights ship in a format with no readable header.

Memory by precision

Everything resident, at the model's native output size, on a card large enough that precision is the only constraint.

Precision	Peak VRAM	Quality	What it costs you
BF16	7.55 GB	100%	How the weights are published. No loss, and the largest footprint.
FP8	4.47 GB	97%	Halves the denoiser with a small, usually invisible cost. Needs Ada or newer.
GGUF Q8_0	4.67 GB	98%	Works on any card, unlike FP8. Slightly slower than native precision.
GGUF Q5_K_M	3.59 GB	95%	A middle step when Q8 will not fit.
GGUF Q4_K_M	3.26 GB	91%	The usual way a 12B image model gets onto an 8 GB card. Detail softens.
NF4	3.09 GB	89%	Aggressive 4-bit. Fast to load, noticeably looser on fine detail.

Memory by arrangement

The other lever: the same weights at the same precision, moved around differently. On this model quantising is the stronger lever: the denoiser is 100% of the pipeline and stays resident whatever you rearrange.

Arrangement	Peak VRAM	Time	How it works
Everything resident	7.55 GB	5 min 2 s	All components stay on the GPU. Fastest, and needs the most memory.
Text encoder on CPU	7.55 GB	5 min 2 s	The prompt is encoded once on the processor, so the encoder never touches the GPU at all. Standard practice on video models, whose encoders are often larger than their denoisers.
Component offload	7.55 GB	5 min 2 s	One component on the GPU at a time. The text encoder runs, then makes way for the denoiser. Costs a few seconds per generation.
Sequential offload	2.14 GB	5 min 2 s	Weights stream layer by layer from system RAM. Runs almost anything on almost anything, and is many times slower.

Which hardware runs MusicGen Large

118 of 118 consumer devices run it in some arrangement. Each row shows the fastest arrangement that fits on that device.

Device	Memory	Verdict	Arrangement	Peak	Time
GeForce RTX 5090 NVIDIA	32 GB	Slow	BF16 everything resident	7.55 GB	1.8x real time
RTX PRO 6000 Blackwell NVIDIA	96 GB	Slow	BF16 everything resident	7.55 GB	1.8x real time
GeForce RTX 3070 Ti NVIDIA	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	1.1x real time
GeForce RTX 4090 NVIDIA	24 GB	Slow	BF16 everything resident	7.55 GB	1x real time
GeForce RTX 3090 Ti NVIDIA	24 GB	Slow	BF16 everything resident	7.55 GB	1x real time
Arc A750 Intel	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	1x real time
GeForce RTX 5080 NVIDIA	16 GB	Slow	BF16 everything resident	7.55 GB	1x real time
RTX 6000 Ada Generation NVIDIA	48 GB	Slow	BF16 everything resident	7.55 GB	1x real time
GeForce RTX 3090 NVIDIA	24 GB	Slow	BF16 everything resident	7.55 GB	0.9x real time
GeForce RTX 3080 Ti NVIDIA	12 GB	Slow	BF16 everything resident	7.55 GB	0.9x real time
GeForce RTX 3080 12GB NVIDIA	12 GB	Slow	BF16 everything resident	7.55 GB	0.9x real time
GeForce RTX 5070 Ti NVIDIA	16 GB	Slow	BF16 everything resident	7.55 GB	0.9x real time
GeForce RTX 5090 Laptop NVIDIA	24 GB	Slow	BF16 everything resident	7.55 GB	0.9x real time
GeForce RTX 5060 Ti 8GB NVIDIA	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	0.8x real time
GeForce RTX 5060 NVIDIA	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	0.8x real time
GeForce RTX 3070 NVIDIA	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	0.8x real time
GeForce RTX 3060 Ti NVIDIA	8 GB	Slow	GGUF Q8_0 everything resident	4.67 GB	0.8x real time
Radeon RX 7900 XTX AMD	24 GB	Slow	BF16 everything resident	7.55 GB	0.8x real time
Apple M3 Ultra 96GB Apple	96 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M3 Ultra 256GB Apple	256 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M3 Ultra 512GB Apple	512 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
GeForce RTX 5080 Laptop NVIDIA	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.8x real time
RTX A6000 NVIDIA	48 GB	Barely usable	BF16 everything resident	7.55 GB	0.8x real time
RTX A5000 NVIDIA	24 GB	Barely usable	BF16 everything resident	7.55 GB	0.8x real time
Apple M1 Ultra 64GB Apple	64 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M1 Ultra 128GB Apple	128 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M2 Ultra 64GB Apple	64 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M2 Ultra 128GB Apple	128 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
Apple M2 Ultra 192GB Apple	192 GB	Barely usable	BF16 everything resident	7.05 GB	0.8x real time
GeForce RTX 3080 10GB NVIDIA	10 GB	Barely usable	BF16 everything resident	7.55 GB	0.8x real time
Radeon PRO W7900 AMD	48 GB	Barely usable	BF16 everything resident	7.55 GB	0.8x real time
GeForce RTX 4080 SUPER NVIDIA	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.7x real time
GeForce RTX 4080 NVIDIA	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.7x real time
Radeon RX 7900 XT AMD	20 GB	Barely usable	BF16 everything resident	7.55 GB	0.7x real time
GeForce RTX 5070 NVIDIA	12 GB	Barely usable	BF16 everything resident	7.55 GB	0.7x real time
GeForce RTX 4070 Ti SUPER NVIDIA	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.7x real time
GeForce GTX 1660 SUPER NVIDIA	6 GB	Barely usable	GGUF Q8_0 everything resident	4.67 GB	0.6x real time
GeForce RTX 2080 Ti NVIDIA	11 GB	Barely usable	BF16 everything resident	7.55 GB	0.6x real time
GeForce RTX 4090 Laptop NVIDIA	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.6x real time
Radeon RX 9070 XT AMD	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.6x real time
Radeon RX 9070 AMD	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.6x real time
Arc A770 16GB Intel	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.6x real time
Radeon RX 7800 XT AMD	16 GB	Barely usable	BF16 everything resident	7.55 GB	0.5x real time
GeForce RTX 4060 Ti 8GB NVIDIA	8 GB	Barely usable	GGUF Q8_0 everything resident	4.67 GB	0.5x real time
Apple M4 Max 36GB Apple	36 GB	Barely usable	BF16 everything resident	7.05 GB	0.5x real time

Source: facebook/musicgen-large . Downloaded 0.1 million times in the last month. See how we calculate, or browse every audio model.