How much VRAM does Kokoro 82M need?

At BF16 with everything resident, Kokoro 82M peaks at about 1.55 GB. Offloading components one at a time brings that to 1.55 GB, and quantising the weights to 4-bit brings it to 1.44 GB.

What is the cheapest GPU that runs Kokoro 82M?

The GeForce GTX 1650 at around $100, running it at BF16 with everything resident, taking about 37 s per generation.

Kokoro · Autoregressive · Text to speech

Kokoro 82M

Kokoro 82M is a single network of 0.082 billion parameters, of which 0.082 billion do the actual generating. It produces the whole utterance in parallel rather than token by token, which is why models this small can run many times faster than real time.

Eighty-two million parameters and better than models twenty times its size. Runs on a phone.

Whole pipeline0.082B

Denoiser0.082B

Download0.3 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	0.082B	F32	0.31 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	1.55 GB	100%	How the weights are published. No loss, and the largest footprint.
FP8	1.48 GB	97%	Halves the denoiser with a small, usually invisible cost. Needs Ada or newer.
GGUF Q8_0	1.48 GB	98%	Works on any card, unlike FP8. Slightly slower than native precision.
GGUF Q5_K_M	1.45 GB	95%	A middle step when Q8 will not fit.
GGUF Q4_K_M	1.45 GB	91%	The usual way a 12B image model gets onto an 8 GB card. Detail softens.
NF4	1.44 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	1.55 GB	700 ms	All components stay on the GPU. Fastest, and needs the most memory.
Text encoder on CPU	1.55 GB	700 ms	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	1.55 GB	700 ms	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	1.42 GB	700 ms	Weights stream layer by layer from system RAM. Runs almost anything on almost anything, and is many times slower.

Which hardware runs Kokoro 82M

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
RTX PRO 6000 Blackwell NVIDIA	96 GB	Comfortable	BF16 everything resident	1.55 GB	137.4x real time
GeForce RTX 5090 NVIDIA	32 GB	Comfortable	BF16 everything resident	1.55 GB	114.5x real time
RTX 6000 Ada Generation NVIDIA	48 GB	Comfortable	BF16 everything resident	1.55 GB	100x real time
GeForce RTX 4090 NVIDIA	24 GB	Comfortable	BF16 everything resident	1.55 GB	90.1x real time
GeForce RTX 5090 Laptop NVIDIA	24 GB	Comfortable	BF16 everything resident	1.55 GB	71x real time
NVIDIA DGX Spark 128GB NVIDIA	128 GB	Comfortable	BF16 everything resident	1.55 GB	68.3x real time
GeForce RTX 5080 NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	61.5x real time
GeForce RTX 4080 SUPER NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	56.8x real time
GeForce RTX 4080 NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	53.3x real time
GeForce RTX 4090 Laptop NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	54.6x real time
GeForce RTX 5070 Ti NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	48.4x real time
GeForce RTX 4070 Ti SUPER NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	48.1x real time
GeForce RTX 5080 Laptop NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	51.9x real time
GeForce RTX 4070 Ti NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	43.7x real time
GeForce RTX 3090 Ti NVIDIA	24 GB	Comfortable	BF16 everything resident	1.55 GB	43.7x real time
RTX A6000 NVIDIA	48 GB	Comfortable	BF16 everything resident	1.55 GB	42.3x real time
GeForce RTX 3090 NVIDIA	24 GB	Comfortable	BF16 everything resident	1.55 GB	38.8x real time
GeForce RTX 4080 Laptop NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	40.4x real time
GeForce RTX 4070 SUPER NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	38.5x real time
GeForce RTX 3080 Ti NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	37.2x real time
GeForce RTX 5070 NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	33.6x real time
GeForce RTX 3080 12GB NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	32.5x real time
GeForce RTX 3080 10GB NVIDIA	10 GB	Comfortable	BF16 everything resident	1.55 GB	32.5x real time
Radeon RX 7900 XTX AMD	24 GB	Comfortable	BF16 everything resident	1.55 GB	33.6x real time
Radeon PRO W7900 AMD	48 GB	Comfortable	BF16 everything resident	1.55 GB	33.3x real time
GeForce RTX 4070 NVIDIA	12 GB	Comfortable	BF16 everything resident	1.55 GB	31.7x real time
GeForce RTX 2080 Ti NVIDIA	11 GB	Comfortable	BF16 everything resident	1.55 GB	29.5x real time
RTX A5000 NVIDIA	24 GB	Comfortable	BF16 everything resident	1.55 GB	30.3x real time
Radeon RX 7900 XT AMD	20 GB	Comfortable	BF16 everything resident	1.55 GB	28.1x real time
GeForce RTX 5060 Ti 16GB NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	26.2x real time
GeForce RTX 5060 Ti 8GB NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	26.2x real time
Radeon RX 9070 XT AMD	16 GB	Comfortable	BF16 everything resident	1.55 GB	26.5x real time
GeForce RTX 4070 Laptop NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	25.4x real time
Radeon RX 7900 GRE AMD	16 GB	Comfortable	BF16 everything resident	1.55 GB	25.1x real time
GeForce RTX 4060 Ti 16GB NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	24x real time
GeForce RTX 4060 Ti 8GB NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	24x real time
GeForce RTX 3070 Ti NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	23.8x real time
Jetson AGX Orin 32GB NVIDIA	32 GB	Comfortable	BF16 everything resident	1.55 GB	23.2x real time
Jetson AGX Orin 64GB NVIDIA	64 GB	Comfortable	BF16 everything resident	1.55 GB	23.2x real time
GeForce RTX 3070 NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	22.1x real time
Radeon RX 9070 AMD	16 GB	Comfortable	BF16 everything resident	1.55 GB	21.3x real time
GeForce RTX 5060 NVIDIA	8 GB	Comfortable	BF16 everything resident	1.55 GB	21x real time
RTX A4000 NVIDIA	16 GB	Comfortable	BF16 everything resident	1.55 GB	21x real time
Radeon RX 7800 XT AMD	16 GB	Comfortable	BF16 everything resident	1.55 GB	20.5x real time
Radeon RX 7700 XT AMD	12 GB	Comfortable	BF16 everything resident	1.55 GB	19.1x real time

Source: hexgrad/Kokoro-82M . Downloaded 11.7 million times in the last month. See how we calculate, or browse every audio model.