README.md

---
title: Nemotron-Labs-Audex
emoji: 🎧
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Unified audio-text intelligence
python_version: "3.10"
startup_duration_timeout: 1h
---

# Nemotron-Labs-Audex

A Gradio demo for
[`nvidia/Nemotron-Labs-Audex-30B-A3B`](https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B)
and [`nvidia/Nemotron-Labs-Audex-2B`](https://huggingface.co/nvidia/Nemotron-Labs-Audex-2B).
The model selector defaults to the 30B-A3B and can switch to 2B model.

The demo includes **audio understanding, speech recognition, speech translation, text reasoning, text-to-speech, and speech-to-speech**.

* Audio inputs support up to 15 minutes.
* Text generation defaults to 1,024 tokens and supports up to 4,096. `Max new tokens` sets a total cap shared by reasoning and the final answer. Reasoning has no separate token cap by default for most tasks.
* TTS defaults to 256 speech tokens and supports up to 512.
* Speech-to-speech uses a 3,584-token reasoning budget within a 4,096-token total output limit.

The hosted ZeroGPU runtime uses a prebuilt `mamba-ssm==2.3.2.post1` Blackwell wheel.

## Run locally

```bash
git clone https://huggingface.co/spaces/nvidia/Nemotron-Labs-Audex
cd Nemotron-Labs-Audex
bash setup_local.sh
bash run_local.sh
```

`setup_local.sh` creates an isolated `.venv` and installs PyTorch when needed.
Use `AUDEX_TORCH_PACKAGE` or standard pip index environment variables to select a platform-specific PyTorch build.

`run_local.sh` detects the selected GPU architecture and creates a matching local CUDA-extension cache.