How much VRAM does orpheus-3b-0.1-pretrained need?

Q4_K_M is exactly 2,363,760,640 bytes (2.20 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is orpheus-3b-0.1-pretrained's KV cache?

3.50 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

canopylabs · audio tts

orpheus-3b-0.1-pretrained

canopylabs/orpheus-3b-0.1-pretrained

orpheus-3b-0.1-pretrained at Q4_K_M is exactly 2,363,760,640 bytes (2.20 GiB / 2.36 GB) — an effective 4.999 bits per weight, not the nominal 4. Its KV cache at 32K is 3.50 GiB.

From the file· summed from 1 file(s)From the file· KV from mirror (mirror:unsloth/orpheus-3b-0.1-pretrained)

Parameters

3.8B

Architecture

llama

28 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XXS	1.40 GiB	1,498,995,616	3.170	—	Mungert
IQ2_XS	1.47 GiB	1,579,113,376	3.339	—	Mungert
Q2_K	1.49 GiB	1,595,321,920	3.374	—	QuantFactory
IQ2_S	1.50 GiB	1,615,240,096	3.416	—	Mungert
IQ2_M	1.55 GiB	1,665,571,744	3.522	—	Mungert
Q2_K_S	1.58 GiB	1,694,817,184	3.584	—	Mungert
IQ3_XXS	1.67 GiB	1,796,119,456	3.798	—	Mungert
Q2_K_M	1.70 GiB	1,822,880,224	3.855	—	Mungert
Q3_K_S	1.70 GiB	1,823,200,480	3.856	—	QuantFactory
IQ3_XS	1.71 GiB	1,834,654,624	3.880	—	Mungert
Q3_K_M	1.83 GiB	1,967,510,752	4.161	—	QuantFactory
IQ3_M	1.84 GiB	1,975,032,736	4.177	—	Mungert
IQ3_S	1.84 GiB	1,975,032,736	4.177	—	Mungert
Q3_K_S	1.88 GiB	2,017,352,608	4.266	—	Mungert
Q2_K_L	1.92 GiB	2,056,406,944	4.349	—	Mungert
Q3_K_L	1.95 GiB	2,095,699,168	4.432	—	QuantFactory
Q4_0	1.99 GiB	2,137,274,272	4.520	—	Mungert
Q3_K_M	2.00 GiB	2,145,415,648	4.537	—	Mungert
IQ4_XS	2.01 GiB	2,158,424,128	4.564	—	Mungert
IQ4_NL	2.11 GiB	2,261,570,752	4.783	—	Mungert
Q4_0	2.11 GiB	2,261,573,632	4.783	—	QuantFactory
Q4_K_S	2.12 GiB	2,272,583,680	4.806	—	QuantFactory
Q4_K_M	2.20 GiB	2,363,760,640	4.999	—	QuantFactory
Q4_1	2.21 GiB	2,373,700,000	5.020	—	Mungert
Q3_K_L	2.22 GiB	2,378,942,368	5.031	—	Mungert
Q4_K_S	2.25 GiB	2,414,572,000	5.106	—	Mungert
Q4_1	2.30 GiB	2,467,866,880	5.219	—	QuantFactory
Q4_K_M	2.30 GiB	2,472,571,360	5.229	—	Mungert
Q5_0	2.43 GiB	2,610,125,728	5.520	—	Mungert
Q5_0	2.49 GiB	2,674,160,128	5.655	—	QuantFactory
Q5_K_S	2.49 GiB	2,674,160,128	5.655	—	QuantFactory
Q4_K_L	2.52 GiB	2,706,098,080	5.723	—	Mungert
Q5_K_M	2.54 GiB	2,726,801,920	5.766	—	QuantFactory
Q5_K_S	2.58 GiB	2,768,269,792	5.854	—	Mungert
Q5_K_M	2.61 GiB	2,798,350,816	5.918	—	Mungert
Q5_1	2.65 GiB	2,846,551,456	6.020	—	Mungert
Q5_1	2.68 GiB	2,880,453,376	6.091	—	QuantFactory
Q5_K_L	2.82 GiB	3,031,877,536	6.412	—	Mungert
Q6_K_M	2.90 GiB	3,112,530,400	6.582	—	Mungert
Q6_K	2.90 GiB	3,112,533,280	6.582	—	QuantFactory

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.44 GiB	0.44 GiB	—	28 / 0 / 0
8,192	0.88 GiB	0.88 GiB	—	28 / 0 / 0
16,384	1.75 GiB	1.75 GiB	—	28 / 0 / 0
32,768	3.50 GiB	3.50 GiB	—	28 / 0 / 0
65,536	7.00 GiB	7.00 GiB	—	28 / 0 / 0
131,072	14.00 GiB	14.00 GiB	—	28 / 0 / 0

Compare with

same modality, comparable size

vs Ace-Step1.5 vs Qwen3-TTS-12Hz-0.6B-Base vs OmniVoice vs orpheus-3b-0.1-ft vs neutts-air vs s2-pro All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 1.98 GiB. The real file is 2.20 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from mirror:unsloth/orpheus-3b-0.1-pretrained

Layers

Attention heads

KV heads

Head dim

128

Hidden size

3072

Vocab

156,939

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does orpheus-3b-0.1-pretrained need?: Q4_K_M is exactly 2,363,760,640 bytes (2.20 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is orpheus-3b-0.1-pretrained's KV cache?: 3.50 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of orpheus-3b-0.1-pretrained should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.