How to Choose the Right Whisper Model for Your Hardware

Whisper ships in several sizes, and picking the right one is a trade-off between speed, accuracy, and how powerful your computer is — not a one-size-fits-all decision.

The trade-off

• tiny — fastest, least accurate. Good for quick drafts or very clear, simple audio.
• base — still fast, a step up in accuracy. A reasonable default on modest hardware.
• small — noticeably more accurate, still runs comfortably on most modern laptops. The largest model available on the free plan.
• medium — strong accuracy, handles accents and moderate background noise well. Needs more RAM and time.
• large — the most accurate Whisper model, best for noisy audio, heavy accents, or professional/legal use where mistakes are costly. Slowest, and benefits most from GPU acceleration.

Matching model to hardware

• 8GB RAM, no dedicated GPU: tiny, base or small will run comfortably. Medium is usable but slower.
• 16GB+ RAM, or GPU acceleration available (Metal on Mac, Vulkan on Windows): medium and large become practical, especially for longer files.

A practical approach

Start with small for everyday transcription. If you're getting mistakes on names, jargon, or noisy recordings, step up to medium or large for that specific file rather than changing your default — ScribeForge lets you pick the model per transcription, not just globally.

Medium and large models require a paid plan; free-plan transcription uses tiny, base or small. See how ScribeForge uses Whisper for more background on the model itself.