Can I run whisper-large-v3 locally?
openai/whisper-large-v3whisper-large-v3 is a seq2seq / translation model with about 1.5B parameters. The practical minimum to run it is roughly 1.8 GB of GPU or system memory, and a GPU with ≥ 2 GB VRAM runs it comfortably. Here's exactly what it needs and how to run it.
Check YOUR exact machinewhisper-large-v3 memory & VRAM requirements
How much memory whisper-large-v3 needs at each quantization level (weights plus ~20% runtime overhead). Lower-bit quantization dramatically reduces VRAM at a small quality cost.
| Precision | Memory needed | Notes |
|---|---|---|
| Full precision (FP16/BF16) | 4.7 GB | Full quality |
| 8-bit (INT8 / Q8) | 2.9 GB | Smaller, slight quality loss |
| 4-bit (Q4_K_M / INT4) | 2.0 GB | Best size/quality trade-off |
| 3-bit (Q3_K) | 1.8 GB | Smaller, slight quality loss |
Will whisper-large-v3 run on your GPU?
Verdicts for common setups — from CPU-only laptops to an RTX 4090 and data-center GPUs. Click a GPU to see everything it can run.
| Your setup | Verdict | Needs | Est. speed | Best way to run |
|---|---|---|---|---|
| No GPU (16 GB RAM) | Workarounds | 4.7 GB | ~13 tok/s | Run on CPU (RAM) with FP16 |
| RTX 4060 (8 GB) | Runs | 4.7 GB | ~55 tok/s | Run on your GPU with FP16 |
| RTX 3060 (12 GB) | Runs | 4.7 GB | ~70 tok/s | Run on your GPU with FP16 |
| RTX 4070 Super (12 GB) | Runs | 4.7 GB | ~93 tok/s | Run on your GPU with FP16 |
| RTX 4080 Super (16 GB) | Runs | 4.7 GB | ~130 tok/s | Run on your GPU with FP16 |
| RTX 4090 (24 GB) | Runs | 4.7 GB | ~160 tok/s | Run on your GPU with FP16 |
| RTX 3090 (24 GB) | Runs | 4.7 GB | ~150 tok/s | Run on your GPU with FP16 |
| Apple M-series (18 GB unified) | Runs | 4.7 GB | ~32 tok/s | Run on your Apple Silicon (unified memory) with FP16 |
| Apple M Max (64 GB unified) | Runs | 4.7 GB | ~77 tok/s | Run on your Apple Silicon (unified memory) with FP16 |
| A100 (80 GB) | Runs | 4.7 GB | ~240 tok/s | Run on your GPU with FP16 |
How to run whisper-large-v3 locally
Recommended method: Run on your GPU with FP16 using vLLM or Transformers.
# Fast production serving on http://localhost:8000/v1
vllm serve openai/whisper-large-v3from transformers import pipeline
pipe = pipeline("text-generation", model="openai/whisper-large-v3", device_map="auto")
print(pipe("Hello", max_new_tokens=50))whisper-large-v3 — frequently asked questions
How much VRAM does whisper-large-v3 need?
whisper-large-v3 needs roughly 2.0 GB of VRAM at 4-bit (Q4) quantization and about 4.7 GB at full FP16 precision, including runtime overhead. Lower-bit quantization trades a little quality for a lot less memory.
Can I run whisper-large-v3 on CPU without a GPU?
Yes — whisper-large-v3 can run on CPU using system RAM (best with a quantized GGUF build via llama.cpp or Ollama), but generation will be noticeably slower than on a GPU.
What's the best way to run whisper-large-v3 locally?
The easiest path is usually vLLM or Transformers. Run on your GPU with FP16. This page's "How to run it" section has copy-paste commands.
What GPU do I need to run whisper-large-v3?
A GPU with at least 2 GB of VRAM runs whisper-large-v3 comfortably at 4-bit quantization, or about 5 GB for full FP16 precision. On Apple Silicon, unified memory of that size works too.