Whisper Large v3

OpenAI · 1.6B · up to 448 context · mit

FITS · 35.3 GB FREE

Runs comfortably on any Apple Silicon Mac and remains the most accurate open transcription model across accents and background noise. Real-time factor on an M-series chip is fast enough for near-live captioning, though the turbo variant trades a little accuracy for more speed.

On a M4 Pro with 48 GB

macOS9.6 GBweights3.1 GBKV cache0.0 GBheadroom35.3 GB

Weights: 3.1 GB
KV cache: 0.0 GBestimated shape
Usable RAM: 38.4 GB
Headroom: 35.3 GB

Generation: 68.7 tok/sest.
Prompt processing: 583 tok/sest.
First token (1K prompt): 1.8 sest.

Decode speed is bandwidth ÷ active weight bytes, at 273 GB/s and 78% efficiency. These are modelled figures, not measurements.

Every quantization, on your Mac

Quant	Weights	+ KV	Verdict	Speed est.	Repo
q4	0.9 GB	0.9 GB	FITS · 37.5 GB FREE	244 tok/s	whisper-large-v3-mlx-4bit
q8	1.6 GB	1.6 GB	FITS · 36.8 GB FREE	129 tok/s	whisper-large-v3-mlx-8bit
bf16	3.1 GB	3.1 GB	FITS · 35.3 GB FREE	68.7 tok/s	whisper-large-v3-mlx

We never host weights. Every link goes to Hugging Face.

Run it

Chat

mlx_lm.chat --model mlx-community/whisper-large-v3-mlx

Serve an OpenAI-compatible endpoint

mlx_lm.server --model mlx-community/whisper-large-v3-mlx --port 8080

How far can you push the context

Which Macs run this

Chip	Smallest RAM that fits	Speed est.
M1	8 GB	15.8 tok/s
M1 Pro	16 GB	49.0 tok/s
M1 Max	32 GB	102 tok/s
M1 Ultra	64 GB	209 tok/s
M2	8 GB	23.5 tok/s
M2 Pro	16 GB	49.7 tok/s
M2 Max	32 GB	103 tok/s
M2 Ultra	64 GB	212 tok/s
M3	8 GB	23.5 tok/s
M3 Pro	18 GB	36.3 tok/s
M3 Max (14-core CPU)	36 GB	75.5 tok/s
M3 Max (16-core CPU)	48 GB	103 tok/s
M3 Ultra	96 GB	219 tok/s
M4	16 GB	28.6 tok/s
M4 Pro	24 GB	68.7 tok/s
M4 Max (14-core CPU)	36 GB	107 tok/s
M4 Max (16-core CPU)	48 GB	144 tok/s
M5	16 GB	37.0 tok/s
M5 Prounverified	24 GB	78.2 tok/s
M5 Maxunverified	36 GB	164 tok/s