Weights Viz for VS Code

Inspect GGUF, SafeTensors, and ONNX model weight layouts without leaving VS Code.

Features

• Open .gguf, .safetensors, .onnx, and .safetensors.index.json files with the Weights Visualization custom editor by default.
• Right-click a supported model or textproto file in the Explorer and choose Visualize Weights to open the viewer explicitly.
• View a Canvas byte map ordered by exact file address, with visible alignment gaps and drag-to-pan navigation.
• Inspect tensor names, shapes, dtypes, exact offsets, lengths, and storage.
• Search file-level metadata, including GGUF model and tokenizer KV entries.
• Sample values from SafeTensors and supported GGUF types on demand.
• Open *.safetensors.index.json to combine model shards automatically.
• Open any standard model-00001-of-000NN.gguf shard to discover and combine its sibling shards automatically.
• Visualize ONNX external-data address spaces inferred from the manifest without requiring the referenced .data files.
• Compare any two GGUF, SafeTensors, or ONNX models side by side. Exact matching tensor names are correlated; changed, added, and removed tensors are highlighted directly in the maps. This compares tensor metadata and encoded layout, not tensor values or raw file bytes.
• Run Weights Viz: Compare Model Weights..., or select the first model from Explorer with Weights Viz: Select Model for Compare and compare from the second model's context menu.
• Run Weights Viz: Open Model Files to pick model files with the native VS Code dialog, or use the Open files button inside the viewer.
• Run Weights Viz: Open Model URL to inspect public remote models with HTTP Range requests.

VS Code webviews cannot open the browser file dialog, so the viewer asks the extension host for the native picker instead of using an HTML file input.

Privacy

• Local model files are parsed on your machine and are never uploaded.
• The extension reads only files you explicitly open, SafeTensors shards referenced by an index, and model URLs you explicitly request.
• Parsed metadata and requested value samples pass only between the extension host and its local VS Code webview.
• The extension contains no telemetry, analytics, advertising, authentication, or secret-storage integration.
• Network requests occur only when you explicitly run Open Model URL. Requests go to that URL (and referenced remote shards), use HTTP Range where supported, and always omit credentials.

See the project README for the full support matrix and remote-server requirements.