--- title: LarQL Vindex Viewer emoji: πŸ”¬ colorFrom: purple colorTo: blue sdk: static pinned: true license: cc-by-nc-4.0 short_description: Interactive viewer for LarQL transformer vindexes --- # LarQL Vindex Viewer Interactive visualization of the SVD-decomposed feature structure of nine open-weight transformers, built from the [Divinci-AI vindex collection](https://huggingface.co/Divinci-AI). **Two views, both live:** - **3D cylinder** (default) β€” each model is a column of layer rings; feature points spiral around each ring colored by circuit stage (broadcast β†’ domain β†’ entity β†’ prediction). - **2D circuit** (`πŸ”Œ 2D Circuit` button or `?view=flat`) β€” layers as horizontal rows with inter-layer edges between top features, network-style. **Compare mode** (`β‡Œ Compare`) renders any two models side-by-side. The default pair (Qwen3-8B vs Bonsai b1.58 1-bit) is the headline visual: organized rings vs scattered dissolution cloud, the same data that drives Paper 1's 1-bit dissolution claim (var@64 β‰ˆ 0.85 for fp16 vs β‰ˆ 0.10 for 1-bit). **Entity search** (`πŸ”` toolbar input or `?q=paris`) β€” type a token, see which features it activates across the model's depth. The Parisβ†’capital match on Gemma 4 E2B is the LarQL Gate-3 result that Post 2 in our blog series unpacks. **Demo** (`β–Ά Demo`) β€” 12-second scripted tour: build β†’ orbit β†’ reveal compare mode β†’ tagline. URL params: `?model=`, `?view=flat`, `?autoplay`, `?q=`. Combine freely. --- ## Background A **vindex** is a transformer's weights decompiled into a queryable feature database. The viewer renders the top-64 SVD feature directions per layer β€” the directions that absorb ~85% of the down-projection matrix's variance in fp16 models and ~10% (near-Marchenko-Pastur random) in 1-bit models. Full details, papers, and the LarQL toolchain: **[huggingface.co/Divinci-AI](https://huggingface.co/Divinci-AI)** Β· [github.com/Divinci-AI](https://github.com/Divinci-AI)