How much VRAM does s2-pro need?

Q4_K_M is exactly 3,566,165,088 bytes (3.32 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

Is s2-pro a mixture-of-experts model?

Yes — 1 experts with 1 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

fishaudio · audio tts · mixture of experts

s2-pro

fishaudio/s2-pro

s2-pro at Q4_K_M is exactly 3,566,165,088 bytes (3.32 GiB / 3.57 GB) — an effective 6.254 bits per weight, not the nominal 4.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

4.6B

total, not active

Architecture

fish-speech

36 layers

Context

—

native (config.json)

License

other

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	2.40 GiB	2,576,474,208	4.518	—	rodrigomt
Q3_K	2.82 GiB	3,029,197,408	5.312	—	rodrigomt
Q4_K_M	3.32 GiB	3,566,165,088	6.254	—	rodrigomt
Q5_K_M	3.75 GiB	4,031,183,968	7.069	—	rodrigomt
Q6_K	4.21 GiB	4,525,266,528	7.936	—	rodrigomt
Q8_0	5.24 GiB	5,630,037,088	9.873	—	rodrigomt
F16	9.23 GiB	9,906,568,704	17.373	—	rodrigomt

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 Fun-CosyVoice3-0.5B-2512 vs VoxCPM2 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 2.39 GiB. The real file is 3.32 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

—

KV heads

—

Head dim

128

Hidden size

—

Vocab

155,776

Sliding window

none

SWA period

—

MLA

Experts

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does s2-pro need?: Q4_K_M is exactly 3,566,165,088 bytes (3.32 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
Is s2-pro a mixture-of-experts model?: Yes — 1 experts, 1 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of s2-pro 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.