Manycore Tech has open-sourced the 3D Gaussian browser Aholo Viewer, which lets any device's browser smoothly load massive 3D scenes with 1 billion+ Gaussian particles. In performance tests, it uses only half the memory of Fei-Fei Li's team's Spark 2.0, renders 3x faster, and has a 10x higher capacity ceiling.

In plain terms: you can browse enormous 3D worlds in your browser as smoothly as scrolling videos.

• Homepage: https://aholojs.dev/zh-CN/
• GitHub: https://github.com/manycoretech/aholo-viewer

What Is Aholo Viewer

3D Gaussian Splatting (3DGS) is a major breakthrough in 3D vision in recent years: it reconstructs real scenes as billions of learnable 3D Gaussian ellipsoids for real-time rendering. But the problem has always been stuck at the "last hundred meters": how do you let ordinary people view them smoothly in a browser?

Aholo Viewer solves exactly this problem. It's a pure-Web 3DGS rendering engine — no software to install; open a browser and you can load and roam massive 3D scenes.

Core Performance Numbers

For a scene with 300 million Gaussian points, Aholo Viewer's desktop test data:

Technical Approach Comparison

There are currently two mainstream LOD (Level of Detail) organizations for Web 3DGS rendering:

Spark 2.0 uses a Splat-based LOD Tree: merging bottom-up from the granularity of individual Gaussian splats. The upside is layer-by-layer detail loading; the downsides are heavy memory and VRAM overhead and weaker scalability.

Aholo Viewer uses a Chunk-based LOD Tree: the raw data is first cut into N chunks (data blocks), LODs at different levels are then generated per chunk, and level switching happens at chunk granularity at runtime.

The Chunk-based approach's two key advantages:

• More controllable overhead: level switching happens per chunk, so memory scheduling works at a coarser granularity with better cache hits
• Better scalability: chunks are clean data boundaries, making it easier to scale up to city-scale and block-scale scenes in the future

Rendering Pipeline Optimizations

Aholo Viewer applies multiple optimizations at the rendering-pipeline level:

1. Multi-precision data structures reduce VRAM usage
2. Cache pre-computation and on-demand passes compress per-frame GPU overhead
3. Morton Sort and detail culling improve data access efficiency

Stack these optimizations together and they land on a set of numbers users can feel: half the memory, loading 1x faster, rendering 3x faster, and a capacity ceiling 10x higher.

Quick Start

Aholo Viewer is an open-source project and simple to use:

1. Visit the GitHub repository to get the code
2. Deploy locally or to a server following the README
3. Import your 3DGS scene data (.ply / .splat formats supported)
4. Open it in a browser and start exploring

You can also visit the online demo page directly to try it out.

Application Scenarios

• E-commerce product display: products go from static images to rotatable 3D scenes
• Online home and car tours: users roam real spaces right in the browser
• Online museum exhibitions: 3D interactive experiences blend into everyday browsing
• Digital twins: city-scale and building-scale 3D scenes rendered in real time in the browser
• UGC 3D content: ordinary users can shoot with a phone and generate shareable 3D scenes

License

Aholo Viewer is open-sourced on GitHub; see the repository for the specific license information. The project and Spark 2.0 (from Fei-Fei Li's World Labs) take different technical routes, together pushing 3D content toward the Web.