Install only what you need¶
A plain install gives you hold-to-talk dictation and nothing that dictation does not need. Every heavier capability — a camera, a neural voice, speaker embeddings — is an optional dependency that is fetched only when you turn that capability on.
That design has one failure mode, and it is the one this page exists to close: a capability can be one word to type and gigabytes to download. So the price is printed where you choose, not only where you pay.
See what something costs before you turn it on¶
$ yazses features
│ NAME TOGGLE NAME DOWNLOAD ADVICE
│ ○ off Voice-activity overlay overlay ~268 MB optional
│ ○ off Read-Back Loop read-back ~352 MB optional
│ ○ off Glance-Type (camera) gaze ~154 MB experimental — not advised yet
│ ○ off Cocktail Filter (voice focus) cocktail ~3.4 GB experimental — not advised yet
│ ○ off Dictation Reflow reflow optional
DOWNLOAD is what a fresh install fetches in total — the whole dependency closure, not just the top-level package, plus any model files the capability pulls down the first time it runs. cocktail reads as one small feature and resolves to speechbrain, which pulls PyTorch and the NVIDIA CUDA stack. read-back is 12 MB of packages and a 340 MB voice model, and the column has to say 352 MB or it is answering a question about pip rather than about your disk.
The two arrive at different times, which is why yazses features enable names them separately: the packages land while you wait on the command, the model on the first run afterwards.
Engine models
stt-parakeet's figure (~604 MB) includes its ~600 MB model. stt-moonshine shows + because its model size is not known, so expect a further download on first use.
A blank cell means there is nothing to download. Most capabilities are blank: reflow, undo, commands and the rest are pure logic that ships in the base install.
If a row reads ◌ set rather than ● ON or ○ off, your config turns that capability on and nothing in this build reads it — it costs you nothing and does nothing. yazses features disable <name> clears the key. See yazses features.
Two numbers, two questions
The catalogue quotes the full size, because it is a price list read by anyone. yazses features enable <name> quotes what is missing on your machine, which is usually smaller and is the number you actually pay. If you already have PyTorch for something else, enable will say so.
Sizes are read from a table that ships inside the package. Listing capabilities never touches the network (ADR-011).
The tiers¶
The ADVICE column says how much you should want each one:
| Tier | Meaning |
|---|---|
core | dictation itself — not optional, not removable |
recommended (on by default) | a fresh install enables these; they are small or free |
optional | useful, off until you ask |
experimental — not advised yet | works, but has a known sharp edge; enable refuses without --force |
planned — designed, not yet wired | in the catalogue so it is honest, not yet callable |
Nothing in the recommended tier is expensive except the tray (~268 MB of Qt, shared with the optional overlay), which gives you the on-screen state that makes dictation legible. The overlay has been off by default since v2.41.0. On a headless or memory-tight machine, turn both off and the Qt dependency is never fetched:
Turning one on¶
$ yazses features enable read-back
⚠ Large download — this downloads ~12 MB (7 packages), plus ~340 MB of model
files on first use.
Ctrl-C now to stop. `--no-install` prints the packages instead of fetching them.
Above ~250 MB, packages and model files counted together, the warning is loud and tells you how to stop it. It is a warning rather than a prompt on purpose: a prompt you have to answer is one more thing to click through, and knowing mid-download still beats knowing afterwards.
To see the packages without fetching anything (this still switches the capability on; yazses features disable read-back undoes it):
Switched on, but nothing installed (◐)¶
A capability turned on with --no-install, by editing config.toml, or before an update rebuilt the environment shows as ◐ ON in yazses features: on, but dormant, because its packages are missing. To see every one of them and what each needs, then install them all after one question:
yazses features repair --dry-run # list only; installs nothing
yazses features repair # install them, then a summary line per feature
yazses restart
It installs into whatever YazSes was installed with — the uv tool receipt, pipx inject, or your venv — so the next update keeps them. A snap or a frozen .app/.exe cannot take extra packages; repair names the reason and exits non-zero instead of trying.
Keeping an install small¶
- Ask what it costs first —
yazses features --category accessor--tier optnarrows the table to the group you are considering. - Know what an engine really costs.
stt-parakeetis ~4 MB of packages on top of what you have plus a ~600 MB model on first use, and it is more accurate thanwhisper-large-v3on English; see choosing a model for the memory side of that decision. - Turn things back off.
yazses features disable <name>stops loading the capability. It does not uninstall the packages —uv pip uninstalldoes, if you want the disk back. - Start from nothing.
yazses features resetreturns every capability to what a fresh install ships with.
See also¶
- Choosing a model on a low-RAM machine — resident memory, not download size
- Running fully air-gapped — what to fetch before you disconnect
- CPU use and battery on a laptop