Every Whisper model, one honest table.

For most machines: ggml-medium-q5_0.bin (539 MB) balances accuracy and speed; on 8 GB RAM or less use ggml-base-q5_1.bin or ggml-small-q5_1.bin; if you have 16 GB+ and patience, ggml-large-v3-turbo-q5_0.bin gives near-best accuracy at a fraction of large-v3's cost. English-only audio can use the .en variants, which are slightly better at the same size.

Quantization level: weights stored at ~5 or 8 bits instead of 16. q5 files are roughly a third the size of the full model with a small accuracy cost; q8 halves the size with almost none. For speech, q5 quantization is usually indistinguishable on clean audio.

Yes — every link resolves directly to the official Hugging Face repository for each project (ggerganov/whisper.cpp, openai, Systran, distil-whisper). We host nothing ourselves; the SHA-256 checksums shown are from the repositories' own file metadata so you can verify what you downloaded.

Run shasum -a 256 <file> (macOS/Linux) or certutil -hashfile <file> SHA256 (Windows) and compare against the checksum in the table. Hover the truncated hash to see the full value.

ggml/gguf .bin files are for whisper.cpp (CPU/Metal, no Python). safetensors are the original PyTorch weights for transformers/openai-whisper. model.bin under Systran repos is CTranslate2 for faster-whisper. They contain the same learned weights converted for different runtimes — pick the format your tool expects.

Only if you want to run Whisper yourself. If you just need a transcript, that is literally our product: upload the file, get speaker-labelled text with word timestamps in about a minute per half hour — no model downloads, no VRAM, from 36¢ per audio hour.