Visualize scientific data, ML model weights, audio, and more — right inside VS Code.
Optimized for Remote SSH. No file transfer, no size limit.
Supported Formats
| Format | Extensions | How it works |
|---|---|---|
| HDF5 | .h5 .hdf5 .hdf .nx .nxs .nc .nc4 + 20 more |
On-demand loading via h5wasm. Runs on the server, only requested data slices are transferred |
| MATLAB | .mat |
v7.3 (HDF5-based): on-demand. v5/v7: full parse via mat-for-js |
| NumPy | .npy |
Standalone files or blobs inside HDF5. Parsed transparently — shows real shape/dtype |
| SafeTensors | .safetensors |
HuggingFace model format. On-demand tensor loading, even for multi-GB files |
| GGUF | .gguf |
llama.cpp model format. Metadata + tensor tree. Non-quantized tensors visualizable |
| PyTorch | .pt .pth .bin .pkl |
Checkpoint files (ZIP + pickle). On-demand tensor loading, no file size limit |
| EEG (Neuroscan) | .cnt |
Binary format. Random-access channel loading |
| EEG (ANT Neuro) | .cnt |
RIFF/RAW3 compressed. Epoch-based decompression |
| Audio blobs | inside HDF5/MAT | .wav .mp3 .flac .ogg .aac named datasets auto-detected |
| JSON blobs | inside HDF5/MAT | .json named datasets shown with syntax highlighting |
Format is auto-detected from file header (magic bytes), not just extension.
v2.2.1: Fixes PyTorch checkpoint metadata floats saved with pickle
BINFLOAT, such as optimizer arguments (args.lr,weight_decay,eps,betas).
Visualizations
- Line plots — 1D datasets with error bars, auxiliary signals, CSV export
- Heatmaps — 2D datasets with colormaps, complex number support, axis controls
- Tables — Matrix view for numerical and compound datasets
- 3D slicing — Navigate slices of higher-dimensional datasets
- Audio player — Waveform + spectrogram (Linear/Log/Mel), playback with seek/volume/filters
- JSON viewer — Syntax-highlighted, auto-formatted, with copy and word wrap
- ML model browser — Hierarchical tensor tree from dot-separated names (e.g.
model.layers.0.weight)
Remote SSH Optimization
The original vscode-h5web transfers the entire file to the local browser for parsing. A 500 MB file = 500 MB network transfer. Files over 2 GB can't open at all.
H5 Viewer runs the parser on the remote server and sends only the data you're looking at:
| Original H5Web | H5 Viewer | |
|---|---|---|
| Open a 500 MB file | Transfer 500 MB | Transfer ~10 KB metadata |
| File size limit | 2 GB | None |
| Switching datasets | Instant (in memory) | RPC request (~ms) |
Installation
From VS Code Marketplace (Recommended):
Install directly from the VS Code Marketplace, or search "H5 Viewer" in the Extensions panel (Ctrl+Shift+X).
From GitHub Release:
- Go to Releases
- Download the
.vsixfile Ctrl+Shift+P→Extensions: Install from VSIX...
From source:
git clone https://github.com/NieeiM/H5-Viewer.git
cd H5-Viewer
pnpm install && pnpm build
pnpm dlx @vscode/vsce package --no-dependencies --allow-missing-repository
All Extensions
Default viewer for: .h5, .hdf, .hdf5, .hf5, .mat, .cnt, .npy, .safetensors, .gguf, .pt, .pth, .bin, .pkl, .nx, .nxs, .nx5, .nexus, .cxi, .nc, .nc4, .loom, .jld2, .h5ebsd, .edaxh5, .oh5, .dream3d, .geoh5, .h5oina, .h5ad.
For other files: right-click → Open With... → H5 Viewer (any extension).
"workbench.editorAssociations": {
"*.foo": "h5viewer.viewer"
}
Experimental Features
The following are functional but not thoroughly tested:
- MATLAB .mat (v5/v7/v7.3)
- EEG .cnt (Neuroscan + ANT Neuro)
- Audio playback and spectrogram
- SafeTensors, GGUF, and PyTorch (.pt/.pth/.bin/.pkl) model files
- NPY and JSON embedded datasets
Platform Support
Single package works on x86_64 and ARM64 (Linux, macOS, Windows). All parsers are WebAssembly or pure JavaScript — no native binaries.
Credits
Built on:
- H5Web and h5wasm by ESRF
- Audio visualization adapted from vscode-audio-preview by sukumo28 (MIT)
- ANT Neuro CNT RAW3 decompression ported from libeep (LGPL-3.0)
License
GPL-3.0