Whisper Model Inventory

The Rule

openai-whisper caches downloaded model weights at ~/.cache/whisper/ as .pt files. Each model is one file. The whisper CLI and the Python whisper.load_model(name) both read from this directory, so whatever is present is immediately usable — no install step beyond the initial download.

What is downloaded where

Audited 2026-04-19. Paths are ~/.cache/whisper/<name>.pt on each machine.

Model	Size	MacBook Pro	Mac Studio	Notes
tiny	72 MB	Yes	No	Fastest, low accuracy. Debugging only.
base	139 MB	Yes	Yes	Decent default for English word timestamps.
small	461 MB	Yes	Yes	Better accuracy, still fast on M-series.
medium	1.4 GB	Yes	Yes	English near-human quality; slower.
large-v3-turbo	1.5 GB	Yes	Yes	Whisper Turbo — fastest large-quality model. Preferred for production transcription.
large-v3	2.9 GB	Yes	No	Highest accuracy, slowest. Rarely needed over turbo.

Total cache size: MacBook Pro ~6.5 GB, Mac Studio ~3.5 GB.

Turbo (large-v3-turbo)

Released by OpenAI October 2024. Same encoder as large-v3, much smaller decoder trained to match large-v3 output. Roughly 8x faster than large-v3 at near-identical quality for English. This is the right default when the task needs real-quality transcription (word timestamps, long-form audio, noisy environments).

In code: whisper.load_model("large-v3-turbo") or whisper.load_model("turbo") (alias).

Install / verify

Whisper itself is a Python package:

pip show openai-whisper    # confirm installed
which whisper              # confirm CLI on PATH

MacBook Pro: installed, CLI at ~/.local/bin/whisper. Mac Studio: models are cached but the whisper CLI is not on PATH — Python import-and-use still works if a project imports whisper directly, but the CLI is not wired up. Re-install if CLI usage is needed: pip install --user openai-whisper.

Downloading a missing model

First use of a model auto-downloads it from OpenAI’s CDN to ~/.cache/whisper/. Either trigger it from Python:

import whisper
whisper.load_model("large-v3-turbo")   # downloads if missing, otherwise loads from cache

…or via the CLI:

whisper --model large-v3-turbo some-audio.mp3

Download happens once per model per machine. No re-download on subsequent runs.

Syncing models across machines

Models are large binary files. They do NOT belong in a git repo. Two valid ways to get a model onto another machine:

1. Let the machine download it on first use — simplest, works offline from that point on.
2. Copy the .pt file over SSH — faster if the other machine already has the bandwidth-heavy download on local network:

   scp ~/.cache/whisper/large-v3-turbo.pt studio:~/.cache/whisper/

Do not symlink between machines or share the cache dir over cloud storage — the files are binary blobs that openai-whisper loads once into memory, and any lazy-loading cloud sync (Google Drive, iCloud Drive) will add startup latency or corrupt the read.

Picking a model for a task

Task	Recommended
Quick word-boundary detection on English voice notes	base
Real transcription of recorded calls / podcasts	large-v3-turbo
Multi-language audio	large-v3-turbo or large-v3
Running on the VPS or low-memory machine	small (fits in 1 GB RAM with headroom)

Rule of thumb: start with base while iterating, upgrade to large-v3-turbo for the final pass. Going straight to large-v3 is rarely worth the 2x slowdown over turbo.

Related

• audio-slow --words mode uses Whisper for editor-precision word boundaries (~/apps/audio-slow/audio_slow.py).
• youtube-auto-commenter and tms-ops use Whisper for transcription of captured audio.