Whisper VRAM requirements
OpenAI publishes approximate VRAM needs for each Whisper model in the official README:
| Model | Parameters | Required VRAM | Relative speed |
|---|---|---|---|
| tiny / tiny.en | 39 M | ~1 GB | ~10x |
| base / base.en | 74 M | ~1 GB | ~7x |
| small / small.en | 244 M | ~2 GB | ~4x |
| medium / medium.en | 769 M | ~5 GB | ~2x |
| turbo | 809 M | ~6 GB | ~8x |
| large | 1550 M | ~10 GB | 1x |
Speeds are relative to large, measured by OpenAI transcribing English on an A100; real speed varies with language and hardware. turbo is an optimized large-v3 with minimal accuracy loss, per OpenAI. The .en models only do English but tend to be more accurate for it, especially tiny.en and base.en. Turbo is not trained for translation: to translate non-English speech into English, OpenAI says to use a multilingual model (tiny, base, small, medium or large) instead.
What fits your card
- 4 GB: tiny, base, small.
- 6 GB: medium (~5 GB). Turbo (~6 GB) is right at the limit, so expect it to be tight.
- 8 GB: medium or turbo. Turbo is the usual pick for transcription: about 4x faster than medium in OpenAI's table and close to large-v3 in accuracy.
- 12 GB and up: large.
These figures are for OpenAI's whisper package. Other runtimes (faster-whisper, whisper.cpp, Transformers with fp16 or 8-bit) can use less; check their own docs.
If you hit out-of-memory
Pick a smaller model or turbo, close other GPU apps (nvidia-smi), and see How to fix "CUDA out of memory". For speech models in our LLM data, see Nemotron 3.5 Transcribe. Running LLMs too? Use the VRAM calculator and What can I run?.
Sources
- OpenAI Whisper README: available models and languages