Fish Speech on Mac: Run Fish Audio S2 Pro Locally

Short answer: Fish Speech is Fish Audio's open text to speech project, and the model you can realistically run on a Mac is Fish Audio S2 Pro, which Murmur ships through an MLX bridge on Apple Silicon and also runs on Intel Macs. The official repository and its installation docs target Linux or WSL on an NVIDIA GPU or CPU, so on a Mac the practical route is a local app that has already converted the weights, not the repo's own commands. This article covers what S2-Pro is, the exact requirements Fish Audio publishes, how the app path works, how the two routes differ, and what the license actually permits.

Every model detail below comes from pages I opened on 2026-10-19: the Fish Speech GitHub README, the Fish Audio S2-Pro Hugging Face card, the installation and requirements page, and the two license texts. Murmur details come from the app's own model catalog. If a number is not on one of those pages, it is not here.

What Fish Speech is today

Fish Audio now ships two weights from the same repository: Drama 3 and S2-Pro. The README lists Drama 3 as a 35B-A3B mixture of experts (3B active per token) and calls it the new-generation flagship for multi-character scenes and long content. S2-Pro is the previous-generation flagship with faster inference. The README's model table puts S2-Pro at 4B parameters, and the Hugging Face card describes a Dual-Autoregressive design with a 4B slow autoregressive backbone plus a 400M fast autoregressive module, roughly 4.5B in total. Some MLX conversion cards round that up to 5B, which is the figure Murmur's catalog uses. Both models share the same command line and API server; the type is read from the checkpoint's config.json.

S2-Pro is the piece most people mean by "Fish Speech" in a local workflow. Per the Hugging Face card it was trained on over 10 million hours of audio across 80+ languages, and it accepts free-form inline [tag] instructions such as [whisper] or [laughing] rather than a fixed tag list, with more than 15,000 unique tags supported. It also does native multi-speaker generation by separating characters with speaker IDs, and, per the README, it reproduces a timbre from 10 to 30 seconds of reference audio. On a single NVIDIA H200 the card reports a real-time factor of 0.195, about 100 ms to first audio, and 3,000+ acoustic tokens per second.

One detail matters before you download anything: the code and the S2-Pro weights do not carry the same license. The repository notice states that the codebase and its associated weights are released under the Fish Audio Community License, while "the S2-Pro weights remain under the Fish Audio Research License." Those two documents grant very different commercial rights, and we will quote them directly later.

Can you run Fish Speech on a Mac?

From the repository, no, not in any documented way. The installation page has a requirements table: Drama 3 needs at least 48 GB of GPU memory (FP8) and runs on "NVIDIA GPU only: no CPU, Apple MPS or AMD ROCm, and no --compile." S2-Pro needs at least 24 GB and runs on an "NVIDIA GPU or CPU," under a system line that reads "Linux, WSL." A separate warning adds that the compile option "is not supported on Windows and macOS." There is no macOS or Apple MPS row anywhere in the requirements, so the repo path on a Mac is CPU-only at best and outside the documented target. The realistic Mac route is an MLX conversion, which is exactly what Murmur ships.

That leads to a simple decision test. Use the table below to pick a lane before you spend an evening on setup.

How Fish Audio S2 Pro runs inside Murmur

Murmur's catalog lists Fish Audio S2 Pro as a "5B dual-AR model with voice cloning and expression tags," running through an MLX bridge rather than the repo's Python stack. It gives two download sizes: about 11 GB in bf16, and about 6.7 GB in an 8-bit variant that Murmur prefers on lower-memory Macs. Murmur's non-Kokoro models ask for at least 16 GB of unified memory, and the app keeps Kokoro available as the lightweight fallback. If you are on a 16 GB machine, the 16 GB Mac guide covers how the app chooses between variants.

The app exposes a specific set of languages for this model: English (US and UK), Japanese, Chinese, Korean, Spanish, Portuguese, Arabic, Russian, French, and German. The Fish Audio model card lists 80+ languages for the model itself, with Japanese, English, and Chinese as Tier 1 and Korean, Spanish, Portuguese, Arabic, Russian, French, and German as Tier 2. Treat the model card as the ceiling and the app's selector as what you can pick today.

The working steps are short:

1. Install Fish Audio S2 Pro from the model list. Murmur downloads the 8-bit variant on a lower-memory Mac and the bf16 variant on a larger one.
2. Add a reference voice. Fish cloning needs at least 3 seconds of audio, 10 to 30 seconds works best, and the app caps a sample at 210 seconds. The 8-bit MLX variant also requires the exact spoken transcript, so record with the words written down or transcribe the take first.
3. Only clone a voice you own or have written permission to use. The Community License states a consent rule of its own, and it is the right habit under any license.
4. Type the script and add expression tags such as [whisper] or [laughing] where the delivery should change. S2-Pro accepts free-form tags, and Murmur exposes the Fish controls plus speed control.
5. Generate and export. Murmur writes standard audio files, so the result drops straight into an editor; the voice cloning guide walks through preparing a clean sample.

Run it from the repo or use a local app

The honest read: pick the repo when you have an NVIDIA Linux box, need the original checkpoints, or want to fine-tune and serve through SGLang. Pick the app when your only hardware is a Mac and you want the model running today. The voice studio versus DIY comparison covers the broader trade-off between maintaining a Python stack and using a packaged app.

The license decides whether you can ship the audio

Two licenses sit on this project, and they are not interchangeable. The Fish Speech source code and the Drama 3 weights fall under the Fish Audio Community License, version 1.0, last updated October 13, 2026. Section III(a) grants a free commercial license "only while Your Annual Revenue is less than USD 500,000." Once your annual revenue reaches or passes that threshold, you must notify Fish Audio in writing within 30 days, the community license continues for a 90-day transition, and after that any commercial use needs "a separate written license agreement from Fish Audio." Section III(c) is stricter: offering a Hosted Service, meaning third parties generate speech from inputs they supply, always requires a separate license regardless of your revenue.

The S2-Pro weights are different. The Hugging Face card marks them under the Fish Audio Research License, last updated March 7, 2026, and the card states plainly: "Research and non-commercial use is permitted free of charge. Commercial use requires a separate license from Fish Audio." The license text is blunter still: "Any use of the Fish Audio Materials or Derivative Works for a Commercial Purpose requires a separate written license agreement from Fish Audio. No commercial rights are granted under this Agreement." So the $500,000 revenue allowance in the Community License applies to the code and the Drama 3 weights, not to the S2-Pro weights you run through an MLX build.

The two documents share some terms. You own the outputs you generate "to the extent permitted by applicable law," and if you redistribute the materials or ship a product that uses them, you must show the "Built with Fish Audio" attribution. The Community License forbids generating speech that imitates the voice of an identifiable natural person without that person's prior consent. The Research License instead requires you to follow Fish Audio's Acceptable Use Policy, which I could not retrieve, so confirm the consent terms there before you rely on either. If you need commercial rights, the license names the address, business@fish.audio.

Practically, a Murmur purchase licenses the app; it does not relicense Fish Audio's weights. If you plan to sell the audio you generate, read the license that ships with the exact weights you use, and contact Fish Audio if you need a commercial agreement. The commercial license comparison and the open model license checklist are useful for auditing a project.

What can go wrong

• The repo has no documented macOS or Apple MPS path. The requirements table lists Linux and WSL, and the docs state the compile option is not supported on Windows or macOS.
• Drama 3 is NVIDIA only: the docs exclude CPU, Apple MPS, and AMD ROCm, and it needs at least 48 GB of GPU memory in FP8.
• Disk and memory planning: S2-Pro is about 11 GB in bf16 or about 6.7 GB in 8-bit inside Murmur, and the repo pulls the main weights plus a codec checkpoint. Make room before you start.
• Python and dependency friction: the repo installs Python 3.12 through conda or uv, and the docs explicitly note pyaudio install errors that may need conda to fix.
• The 8-bit MLX variant needs the exact reference transcript, so a clean recording with known words saves a retry.
• License mismatch: the code and the S2-Pro weights are governed by different licenses, so a permission that applies to one does not automatically apply to the other.
• The docs say Drama 3 switches to float16 on older GPUs without bfloat16 and has not been tested on real hardware there, and for S2-Pro you may need to pass the --half option yourself.

Frequently asked questions