Qwen3-TTS on your computer

Qwen3-TTS is an open speech model from Alibaba's Qwen team (Apache-2.0 licence). Reelcraft can use it for Generate Speech instead of a paid service: it runs on your own computer, costs nothing per request, and no text leaves it.

Reelcraft sets it up for you. There is nothing to start by hand, and nothing is installed or downloaded until you say so.

Set it up

1. Open a Generate Speech stage (or the Speech and audio default) and choose Qwen3-TTS (local) as the provider. Or open Settings → Qwen3-TTS (local).
2. Select Set up Qwen3-TTS. Reelcraft shows what it will install and how big it is, and how much disk space is free, and waits for you to confirm. This installs the software Qwen3-TTS runs on.
3. Watch the progress. Cancel stops it.
4. Select Download on the voice model you want, check its size, and confirm. The two models are shown together: Qwen3-TTS 0.6B (2.5 GB) and Qwen3-TTS 1.7B (4.5 GB).
5. When it finishes, select the model, then pick a voice and settings as usual.

Reelcraft never downloads or installs anything on its own, and a run never starts a download.

It only runs while you use it

The first time you make speech, Reelcraft starts the service (a few seconds, plus loading the model). After about ten minutes with nothing to do, it stops by itself and frees the memory, and starts again for the next request. Under Settings → Qwen3-TTS (local) you can also Stop it, or Reinstall for this hardware, for example after you add a graphics card.

To free the space, select Delete beside a model, in the same place. The software can be removed by deleting its folder (see Where things are kept).

What you need

• Disk space. The software is about 1.2 GB on a Mac (measured). On other computers that is a guess, and a version for an NVIDIA graphics card is larger, so Reelcraft shows an estimate before it installs. The models are 2.5 GB (0.6B) and 4.5 GB (1.7B), downloaded separately.
• Memory. On a graphics card the 0.6B model needs roughly 3 GB of its memory and the 1.7B roughly 5 GB. On the processor alone, roughly 4–5 GB of RAM for the 0.6B and 9 GB or more for the 1.7B. These are estimates from the file sizes, not measurements. If memory is short, use the 0.6B model.
• An internet connection for the first setup (Python packages, a Python itself, and the model).
• uv, if Reelcraft runs directly on your computer rather than in Docker. It installs the software. If it is missing, Reelcraft shows the command for your computer and a Check again button. The Docker image already has it.

Which processor it uses

Reelcraft picks the fastest one it finds and tells you, as Running on: in the setup panel:

Your computer	What it uses
Mac with Apple Silicon (not Docker)	The Apple GPU
NVIDIA graphics card	The graphics card (see Docker for Docker)
Anything else	The processor, which is slower

If the graphics card is out of memory or fails, Reelcraft retries on the processor and says why. The setup installs the version of PyTorch for your computer (this is uv's --torch-backend auto). If the GPU version can't be installed, select Install for CPU instead.

The Apple GPU and the processor have been tested. NVIDIA graphics cards, AMD and Intel cards have not been tested by the Reelcraft team yet.

In Docker

Everything works inside the container, on the processor: the setup is the same, and the software and models are kept in the reelcraft-data volume under /data/qwen-tts.

Use an NVIDIA graphics card

Docker Desktop's Run button has no option for graphics cards, so start Reelcraft from a terminal, with --gpus all. If a container named reelcraft already exists, stop and remove it first (your data stays in the reelcraft-data volume), then run:

docker run -d --name reelcraft --restart unless-stopped --gpus all \
  -p 8080:8080 -v reelcraft-data:/data \
  zohaibkhan97/reelcraft

Then select Reinstall for this hardware under Settings → Qwen3-TTS (local) so the graphics card version is installed.

Windows

You need an up-to-date NVIDIA driver on Windows, and Docker Desktop with the WSL 2 backend. See Docker's guide to GPU support in Docker Desktop.

Linux

You need the NVIDIA driver and the NVIDIA Container Toolkit on your computer. See NVIDIA's install guide. On Docker Engine, also add --add-host=host.docker.internal:host-gateway as in Run Reelcraft.

This has not been tested by the Reelcraft team. If the setup panel still says No GPU is visible inside Docker, the card is not reaching the container.

On a Mac

Docker on a Mac can't use the Mac's GPU, so Qwen3-TTS runs on the processor there, which is slow. For faster speech, run the service on the Mac itself and point Reelcraft at it.

Memory

Docker Desktop limits the memory containers get (Settings → Resources). The 1.7B model on the processor needs more than the 4 GB the install guide asks for. Raise the limit, or use the 0.6B model.

Where things are kept

Locally, the software, the models and a small state.json are in ~/.reelcraft/qwen-tts. Set QWEN_TTS_HOME to use another folder. In Docker they are in /data/qwen-tts in the volume.

They are left out of backups: a restored Reelcraft sets the software up again and downloads the model again, when you ask it to.

Setting	What it does
QWEN_TTS_HOME	Where the software and models are kept
QWEN_TTS_URL	Use a service you run yourself instead (see below)
QWEN_TTS_DEVICE	Force cuda:0, mps or cpu instead of choosing automatically
QWEN_TTS_IDLE_SECONDS	How long a loaded model stays in memory unused (default 600)

Use model files you already have

If you already downloaded a model, point Reelcraft at the folder instead of downloading it again. Under a model that is not downloaded, open Already have the files?, type the folder's full path, and select Check folder. Reelcraft checks that every file is there with the right size, and says which are missing. The files are used where they are: nothing is copied or changed, and Delete only forgets the folder. In Docker, the folder must be shared with the container, and the path is the one inside it.

Use a service you run yourself

You can run the Qwen3-TTS service yourself, on this computer or another one, and tell Reelcraft its address. Then Reelcraft installs and starts nothing, and only uses it. This is also how Reelcraft in Docker on a Mac can use the Mac's GPU.

1. Install uv. From the Reelcraft folder, run pnpm tts. Without a Reelcraft folder, download the services/qwen-tts folder from the repository and run uv run --extra engine qwen-tts-service in it. The first start installs the packages.
2. In Settings → Qwen3-TTS (local), open Use a service you run yourself, enter the address (http://127.0.0.1:8091, or http://host.docker.internal:8091 from Docker), and select Save. Test connection checks it. Reset goes back to letting Reelcraft manage it.

The service has no login and listens on 127.0.0.1, so only this computer can reach it. If a Docker container can't reach it, start it with QWEN_TTS_HOST=0.0.0.0, which opens it to your whole network: do that only on a network you trust.

If it fails

• "Set up isn't available here": this copy of Reelcraft has no service to set up. Run one yourself, as above.
• "Install uv first": run the command shown in a terminal, then select Check again.
• "Not enough disk space": free some space, or delete the other model.
• "The install failed": read the log under the message. If it is about PyTorch or your GPU, select Install for CPU instead.
• "The service stopped unexpectedly": often not enough memory. Try the 0.6B model, or close other apps. Start now tries again.
• "Can't reach the service you run yourself": start it, or Reset the address under Settings → Qwen3-TTS (local).
• The download failed or failed its checksum: select Try again. Files already downloaded are kept.
• A run fails with "isn't downloaded" or "is not set up yet": set it up and download the model, as above.
• Very slow: on the processor a long script takes a while. Try the 0.6B model, a shorter text, or a computer with a graphics card.