Unreal Engine avatars that run on 2GB of VRAM

What it does

GMTalker is a full-stack digital human system built around Unreal Engine 5. It ties together speech recognition, text-to-speech, natural-language understanding, and lip-sync animation to drive 3D cartoon avatars locally. The backend can plug into various LLMs and knowledge bases, while the frontend renders the character and handles voice interruptions and wake-word logic.

The interesting bit

The project is explicitly aimed at “ordinary laptops” and Huawei Ascend NPUs, not just CUDA farms. That hardware flexibility is unusual for UE5-based avatars, which typically assume a beefy GPU and constant cloud connectivity. The Guangming Laboratory also open-sourced the complete UE5 project files, including a government-co-developed mascot named GuangDUNDUN.

Key highlights

• Runs fully offline with a claimed millisecond response time, requiring only 2GB of VRAM for the full stack.
• Single codebase deploys to Windows, Linux, and Android; the client needs no GPU.
• Supports voice interruption, wake-up words, and voice cloning out of the box.
• Backend is model-agnostic, allowing swaps for custom LLMs, agents, or local knowledge bases.
• Ships as a standalone .exe plus Docker images, with ONNX-based ASR and NPU-accelerated TTS.

Caveats

• The license is CC BY-NC 4.0, which forbids commercial use despite the README stating the project is “suitable for… commercial integration.”
• Performance claims such as “millisecond response” are not backed by published benchmarks in the repository.

Verdict

Worth a look if you want a turnkey, offline UE5 avatar for non-commercial demos or interactive agents. Skip it if you need a permissive open-source license or plan to ship a commercial product.