Local AI / Local STT Setup Guide
Short answer
use.fo Local AI / Local STT runs Whisper Large V3 Turbo for speech recognition and Qwen 3.5 0.8B for AI rewriting entirely on your device. It requires macOS Apple Silicon (M1+) or Windows x86/x64 with VRAM > 4GB. After license verification and model download, all processing is offline.
Evidence and updates
Facts checked on July 24, 2026
Reviewed by use.fo product team. Sources below identify whether a fact comes from an official product page, a first-party test, customer evidence, or an independent source.
Steps
- Purchase a Local AI / Local STT license
Go to dashboard.use.fo and purchase a Local AI / Local STT license for $24.99. Each license covers one desktop device. You can transfer the license to a different device once every 6 months.
- Choose the correct local route
Apple Silicon Mac: use the Apple Silicon installer and confirm your chip in Apple menu → About This Mac. Windows: confirm x86/x64 architecture and more than 4 GB available VRAM in Task Manager → Performance → GPU before installing the Local AI route. If either hardware check fails, use Subscription or BYOK instead.
- Install on Apple Silicon Mac
Download usefo-mac-arm64.dmg from use.fo/download, move use.fo to Applications, and open it. Grant microphone access and Accessibility permission so use.fo can paste text into other apps.
- Install on Windows x86/x64
Use the Local AI Microsoft Store listing linked from use.fo/download. Open the app, grant microphone access, and confirm the hardware detection panel accepts your device before downloading the model.
- Sign in to your account
Sign in with the account you used to purchase the Local AI license. The app will verify your license.
- Select Local AI / Local STT in settings
Go to app Settings → Processing Mode → select Local AI / Local STT. The app will detect your hardware and select the appropriate model variant.
- Download the local model
The app will download the Whisper Large V3 Turbo model for STT and the Qwen 3.5 0.8B model for rewriting. Model size is approximately 1–3 GB depending on the quantised variant selected for your hardware. This is a one-time download.
- Allow necessary permissions
Grant microphone access when prompted. On macOS, also grant Accessibility permission so use.fo can paste text into other apps.
- Test your setup
Hold your trigger key, speak a sentence, and release. use.fo should transcribe and rewrite your speech using the local model. The first inference may be slower while the model warms up.
Before you start
Before setting up Local AI, confirm your hardware: • macOS: Apple Silicon (M1, M1 Pro, M1 Max, M1 Ultra, M2, M2 Pro, M2 Max, M2 Ultra, M3, M3 Pro, M3 Max, M4, or later). macOS Intel is not supported. • Windows: x86/x64 CPU architecture with a GPU that has VRAM > 4GB. This includes NVIDIA and AMD discrete GPUs, as well as integrated Intel and AMD graphics if 4GB or more of system memory is allocated as shared GPU memory. Windows ARM is not currently supported.
Troubleshooting
Problem: Model download fails or is very slow
Solution: Ensure you have a stable internet connection for the initial download. The model files are 1–3 GB. If the download fails, restart the app and try again.
Problem: Hardware detection says my device is not supported
Solution: Confirm your macOS device is Apple Silicon (M1 or later). Intel Macs are not supported. On Windows, confirm your GPU has > 4GB VRAM available and that your CPU is x86/x64 architecture.
Problem: Local AI is much slower than expected
Solution: On first run, models may take 10–30 seconds to load into memory. Subsequent uses are faster. If speed remains poor, your VRAM may be at or near the minimum. Closing other GPU-intensive apps can help.
Problem: Transcription quality is lower than expected
Solution: Speak clearly and reduce background noise. Quality can vary with the speaker, vocabulary, microphone, model build, and environment. The published local STT report explains why the current Apple Silicon Whisper route was selected, but it does not claim a universal Local-versus-cloud percentage; test the route that fits your workflow.