v0.9.1

Full Changelog: v0.9.0...v0.9.1

v0.9.0

What's Changed

• feat: support additional model families and streamline role controls by @nvkudva in #26

Full Changelog: v0.8.6...v0.9.0

v0.8.6

Reliability fixes from a full code review.

• A worker that is slow to answer now gets a 504 instead of being restarted mid-job.
• Stop now cancels a retained request while its model downloads or cold-loads.
• Worker crash logs no longer vanish once a log passes 10 MB.
• omni down finishes cleanup within 5 s even with a stream open, so pinned Ollama models are released; the gateway access log is off.
• Inference routes refuse Origin: null (sandboxed iframes on other sites).
• Decide model files are copied atomically, and deleting the model frees its disk.
• TTS input with nothing speakable is a 400, not a 500.

Upgrade: brew update && brew upgrade omni, or rerun install.sh. Then omni restart.

v0.8.3

Changes

• omni init prints a plan before it starts (what downloads, what waits for first use) and says why a speech runtime was skipped, instead of "nothing to warm".
• omni doctor has a new runtimes line for the speech, image and video runtimes.
• A missing media runtime is reported once, with the reinstall command (uv tool install --force "omni-api[media]", or brew reinstall omni), instead of "no worker named stt".
• Install hints and README separate the two steps: install the command and its runtimes, then omni init for the models (image and video download on first use or omni init --media).

v0.8.2

Install experience

• brew install omni now shows uv's progress (resolve, download, build) instead of sitting silent for minutes, and caches wheels in Homebrew's cache dir. A reinstall or upgrade drops from about 2m20s to about 30s.
• install.sh prints an elapsed-time line every 20s during the quiet git clone and build steps, and raises uv's download timeout to 120s so large torch wheels survive slow links.

v0.8.1

A simpler launch film, brighter dark-mode text and a rebuilt playground.

• Launch film re-cut in plain words: "A multi-model server", models instead of roles, one slide per model, and the brew install block in the outro
• Dark mode: brighter text and lifted card surfaces
• Site: simplified landing page that leads with the film
• Playground build refreshed

Install: brew tap nvkudva/omni && brew trust nvkudva/omni && brew install omni, then omni init and omni up.

v0.1.0

First public release.

Install

curl -fsSL https://raw.githubusercontent.com/nvkudva/omni-api/main/install.sh | sh

or with Homebrew:

brew tap nvkudva/omni
brew install omni

Then

omni init     # install Ollama if missing, pull chat, decide and embed
omni up       # gateway on http://127.0.0.1:10000/
omni status

Requires macOS on Apple Silicon: the runtime is MLX, and the stt/tts/image/video workers need Metal. The speech, image and video models download on first use.

PyPI is not a channel yet: the package depends on git-pinned diffusers and ltx revisions that PyPI cannot express, so both the script and the formula install from a source tree.