Turu for Qualcomm AI Hackathon
About
Thuniverse AI has been collaborating closely with Qualcomm in recent months to integrate support for Snapdragon NPU (referred to as NPU) and Qualcomm AI Hub (referred to as AI Hub) into both Kuwa and Turu.
NPU and AI Hub support is already publicly available in Kuwa v0.4.0. Building upon this, Thuniverse AI has continued optimizing the integration and released Turu 25H1-WoS, a version specifically pre-integrated with AI Hub models for participants in the Qualcomm AI Hackathon.
The following table highlights the key differences between Kuwa v0.4.0 and Turu 25H1-WoS:
| Feature | Kuwa v0.4.0 | Turu 25H1-WoS (based-on Kuwa v0.4.0) |
|---|---|---|
| Positioning | Open Source Software | Thuniverse AI Product for Qualcomm AI Hackathon |
| License | MIT License + BSD 3 Clause License | OSS License + Commercial License |
| Default Enabled Model | CPU/GPU-based model (NPU-based model not enabled by default) | NPU-based model (from AI Hub): Image generation: Stable Diffusion v1.5 Text generation: Llama-v3.1-8B-Instruct, Llama-v3.2-3B-Instruct, Phi-v3.5-mini-Instruct, Llama3-TAIDE-8B-LX-Alpha1 Speech recognition: Whisper-Base-En |
| NPU Model Integration | Basic Performance | Optimized Performance |
Download Turu 25H1-WoS
Prerequisite:
- To use the NPU, a device with a Qualcomm Snapdragon X series SoC is required. It is recommended to have at least 16GB of system memory.
- Additionally, installing Turu-25H1-WoS with pre-integrated models requires approximately 12GB of disk space, while installing all models necessitates 25GB of disk space.
You can download Turu-25H1-WoS from our GitHub release (https://github.com/thuniverse-ai/turu-25h1-wos/releases) or Google Drive.
Please note that Turu-25H1-WoS-Setup.exe is an installer that includes pre-integrated Llama-3.2-3B, Whisper, and Stable Diffusion models. You will also need to download all files named Turu-25H1-WoS-Setup-*.bin.
Minimal installation takes about 20 minutes. If you choose to install additional models, the download time may take an extra 20-40 minutes, depending on your network speed.
Training Material
- How to Use Pre-integrated NPU-based Models in Turu 25H1-WoS
- Kuwa Introduction Video (Mandarin)
- Kuwa Tutorials
- Documentation of Turu/Kuwa
- Enabling Qualcomm NPU Support in Kuwa v0.4.0
If you run into any problems, you can reach out to us on GitHub issues or connect with us in the Kuwa community.