| # Space architecture | |
| ## Current release | |
| The Space uses the Hugging Face **static HTML SDK**. The uploaded | |
| `index.html` is self-contained and requires no Python server. | |
| This design provides: | |
| - free static hosting; | |
| - fast startup; | |
| - anonymous public access; | |
| - no secrets; | |
| - no model download; | |
| - interactive graph search, selection, labels, and detail panels; | |
| - downloadable GraphML and JSON. | |
| ## Why the model is not executed inside this Space | |
| The released weights are in Apple MLX format and are validated on Apple | |
| Silicon. Hugging Face Spaces run on Linux CPU or NVIDIA GPU hardware, so | |
| the MLX artifact is not a compatible server-side runtime there. | |
| A future live-classification Space should be a separate deployment using | |
| one of these approaches: | |
| 1. a Linux-compatible source/BF16 or quantized model on GPU hardware; | |
| 2. a remote OpenAI-compatible classification endpoint configured through | |
| Space secrets; | |
| 3. a dedicated Hugging Face Inference Endpoint. | |
| Keeping the static explorer separate prevents the demonstration from | |
| claiming inference that it does not perform. | |