How much VRAM does Whisper large-v3 need? The weights, OpenAI’s planning figure and the measured memory of each program.
Short answer. Plan on 5 to 6 GB of VRAM for Whisper large-v3, and on 10 GB if you do not yet know which program you will use. OpenAI’s README lists about 10 GB for the large model. Measured runs of the same-size large-v2 model on an 8 GB RTX 3070 Ti used 2.9 GB (faster-whisper, 8-bit), 4.1 to 5.0 GB (16-bit in whisper.cpp, faster-whisper, openai/whisper and Transformers) and 6.1 GB (faster-whisper with batch size 8). The 16-bit weights file is 3.1 GB. The turbo model has 809 million parameters, a 1.6 GB file, and OpenAI lists about 6 GB for it.
01
The model itself
Its model card says large-v3 has the same architecture as the earlier large models apart from minor differences, such as 128 mel frequency bins instead of 80 and a new language token. Both have about 1.55 billion parameters, so the memory figures below for large-v2 are a fair guide. OpenAI’s README gives approximate memory requirements for its reference code and does not say what precision they assume.
Model
Parameters
Weights file (16-bit)
VRAM in OpenAI’s table
Whisper large-v3
1.54 billion
3.09 GB
About 10 GB (listed as large)
Whisper large-v3-turbo
0.81 billion
1.62 GB
About 6 GB (listed as turbo)
02
Measured memory by program
These are the faster-whisper project’s measurements on an NVIDIA RTX 3070 Ti with 8 GB (CUDA 12.4). The note in that README says Transformers ran out of memory at any batch size above 1 on that card.
Program (large-v2, 13 minutes of audio, beam size 5)
GPU memory
Compared with the 3.1 GB weights
faster-whisper, 8-bit
2,926 MB
Under the 16-bit file size, because the weights are 8-bit
whisper.cpp with flash attention, 16-bit
4,127 MB
About 1.0 GB above the weights
faster-whisper, 16-bit
4,525 MB
About 1.4 GB above
openai/whisper, 16-bit
4,708 MB
About 1.6 GB above
Transformers (SDPA), 16-bit
4,960 MB
About 1.9 GB above
faster-whisper, 8-bit, batch size 8
4,500 MB
Batching adds about 1.6 GB over the unbatched 8-bit run
faster-whisper, 16-bit, batch size 8
6,090 MB
Batching adds about 1.6 GB over the unbatched run
03
How to plan
One recording at a time: budget 5 to 6 GB for a 16-bit program, or about 3 GB with 8-bit weights.
Several recordings at once, or batching to go faster: add about 1.5 GB for a batch of 8, or more for larger batches.
Sharing the GPU with a chat model: Whisper’s few gigabytes come out of the same pool as the model and its conversations, so count it in the memory budget.
Unsure which program you will run: use OpenAI’s 10 GB for large, or 6 GB for turbo, as the safe figure.
04
Speech recognition and ANVE
ANVE’s meeting minutes work from transcripts in Hindi, English and Hinglish, and every action item must quote the transcript line it came from. This page does not say which speech model ANVE uses, and ANVE has not published speed or memory figures for its launch hardware.
05
Questions people ask
Can Whisper large-v3 run in 4 GB of VRAM?
In the faster-whisper test the large-v2 model used 2.9 GB with 8-bit weights, so it can fit. The effect of 8-bit weights on accuracy was not measured there, so test your recordings.
Does large-v3 need more memory than large-v2?
Little or no more. Both have about 1.55 billion parameters, so the weights are the same size. The measured figures on this page are for large-v2.
Is turbo much lighter than large-v3?
Yes. Turbo has 809 million parameters and a 1.6 GB file, roughly half the size, and OpenAI lists about 6 GB of VRAM against about 10 GB for large.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
