How to Download Whisper Models on Windows
StarWhisper keeps the model step inside the desktop app. You do not need to create a Python environment or manage command-line paths for this workflow.
- Install StarWhisper on 64-bit Windows 10 or Windows 11.
- Open Settings > Models and check which local files are already present.
- Choose a model, start its download, and wait for completion while you have internet access.
- Run a short local test before relying on the model for offline work.
The direct installer can bundle Tiny, Base, and Small, while the Microsoft Store package may require Base or Small to be downloaded. The model manager checks the actual files and lets you download a missing free model.
How to Choose a Model
Model choice is a trade-off between download size, compute, speed, language needs, and how much review your transcript will receive. Start with a free model, then test a Pro model on the same sample if the larger file is worth the extra resources.
When to try Large V3
- Long recordings that will be reviewed before use
- Projects where you can reserve more disk space and processing time
- Cases where you have compared the output on your own hardware and language
When to start smaller
For live dictation and short notes, Small, Base, or Tiny uses less disk space and compute. A larger model can take longer, especially on CPU-only hardware, so compare the workflow before downloading every model.
Language and translation note
Use a multilingual model when your recordings need languages beyond an English-only workflow. The app's model and language settings determine which local path is used; check the selected language before an offline session.
Whisper.cpp in the Desktop Workflow
StarWhisper bundles a native whisper.cpp engine for local inference. The desktop path keeps model selection, hardware choices, and downloads in the app instead of asking you to assemble a separate command-line environment.
- No separate Python environment is required for the app workflow
- Local models are downloaded and checked by the model manager
- CPU, CUDA, and Vulkan paths can be selected by the app when supported
The engine still depends on the selected model and the computer running it. If you need reproducible timing, compare your own sample with the measured Windows benchmark, which documents its hardware and speed-only scope.
Model Storage and Offline Prerequisites
Downloaded models are stored in the app's local data folder. The nominal model sizes in the current source are:
- Tiny: 75 MB
- Base: 142 MB
- Small: 466 MB
- Medium: 1.45 GB
- Large V3 Turbo: 1.55 GB
- Large V3: 2.9 GB
Download only the models you need and leave room for the installer and temporary files. For offline use, finish the installer and model download first, select local mode, and run a short test before disconnecting. The planned offline dictation walkthrough and Notepad dictation walkthrough cover those checks. Free access includes 2,000 words/week; Pro is $10/month or $80/year.
