NVIDIA Audio Flamingo 3 is a large audio–language model for speech, sound, and music understanding. This repository is a Pinokio launcher: it installs dependencies into a project-local virtual environment and runs a Gradio web UI defined in app/app.py.
- Loads the
nvidia/audio-flamingo-3-hfweights (via Hugging Face Hub on first run; large download). - Exposes tabs for single-turn Q&A, transcription, and “think” / long-audio style prompts in the browser.
- Open this app in Pinokio.
- Run Install once (PyTorch via
torch.js, then Python packages fromapp/requirements.txt). - Run Start. When the UI is ready, use Open Web UI from the sidebar, or follow the URL shown in the terminal.
Reset removes the app/env virtual environment. Update runs git pull and upgrades Python dependencies.
- A capable NVIDIA GPU is recommended for practical use; CPU may run but will be slow.
- The Think / Long Audio tab loads a second full copy of the 7B model (plus a LoRA adapter) the first time it is used. On a 16GB GPU the two copies do not fit together, so part of the model is offloaded to system RAM and Think requests run much slower.
- Model use is subject to NVIDIA’s non-commercial research terms; see the About tab in the UI and the model card on Hugging Face.
- If the model weights are gated on Hugging Face, run
hf auth login(or set theHF_TOKENenvironment variable) inside the app's virtual environment before starting, otherwise model loading will fail on first run.
- NVIDIA GPUs use the standard CUDA wheels; AMD GPUs fall back to
torch-directml, which lags behind mainline PyTorch releases — if install fails on an AMD GPU, see the version-pinning note intorch.js. - No CPU-only fallback beyond the default
torch.jsCPU branch is provided; a 7B-parameter model will be very slow without a GPU.
This demo is built for interactive use in the browser through Gradio. There is no separate HTTP API enabled by default.
- Python: Run the app outside Pinokio from the
appfolder with the same virtualenv andpython app.py. You can import and call the underlying model code only by extendingapp/app.pyyourself (not exposed as a stable public API in this launcher). - JavaScript / cURL: Not applicable unless you add a server route or use Gradio’s optional sharing/API features in code; the stock launcher does not start a documented REST surface.
For automation, prefer Pinokio script APIs (install.js, start.js, etc.) or running python app.py under your own process manager.