Whisper Small F16 for Unsloth Studio
Run accurate, fully local speech-to-text dictation in Unsloth Studio. Whisper Small provides better transcription quality than Tiny or Base while remaining practical for everyday local dictation.
Run in Unsloth Studio
- Install or update Unsloth Studio.
- Open Settings > Voice.
- Open the local dictation model picker and select Whisper Small.
- Let Studio download and cache the model.
- Use the microphone button in the chat composer to dictate locally.
The model runs on your device through whisper.cpp. Your recorded audio does not need to be sent to a hosted transcription service.
Model file
whisper-small.bin: native F16 model forwhisper.cpp- Download size: approximately 488 MB
- Best for: improved accuracy with moderate local resource use
whisper.cpp uses a custom GGML binary format for Whisper. The model file is therefore named .bin, not .gguf, even though this repository follows the common -GGUF repository naming convention.
No low-bit quantization was applied. Matrix weights are stored as F16, while tensors that whisper.cpp requires in F32 remain F32.
Manual whisper.cpp usage
whisper-cli -m whisper-small.bin -f audio.wav
Integrity
- Source
model.safetensorsSHA-256:1d7734884874f1a1513ed9aa760a4f8e97aaa02fd6d93a3a85d27b2ae9ca596b - Converted model SHA-256:
cfd85d74dc730828cef4e13ae65898d9dc6f695fd00394f4399f7310e73cf505 - Conversion tool:
ggml-org/whisper.cppcommit080bbbe85230f624f0b52127f1ae1218247989f9
The converted model was loaded by whisper.cpp and passed an end-to-end transcription test.