mlpeel

Minimal-dependency neural network model viewer. Open any supported model file and instantly visualize its architecture as an interactive graph.

npx mlpeel model.onnx

Features

• Minimal dependencies — core viewer uses only Node.js 22+ built-ins; inference requires onnxruntime-node
• Interactive graph — pan, zoom, click nodes to inspect shapes, types, and parameters
• Three formats — ONNX, Safetensors, GGUF
• Auto-detection — format recognized from extension or magic bytes
• JSON mode — dump model structure as JSON for scripting

Supported Formats

| Format | Extension | Use case | |--------|-----------|----------| | ONNX | .onnx | Universal interchange format | | Safetensors | .safetensors | HuggingFace model weights | | GGUF | .gguf | llama.cpp / local LLM inference |

Install

# run directly
npx mlpeel model.onnx

# or install globally
npm install -g mlpeel

Usage

# open model in browser
mlpeel model.onnx

# custom port
mlpeel model.safetensors --port 3000

# don't open browser
mlpeel model.gguf --no-open

# output as JSON (no server)
mlpeel model.onnx --json

Options

mlpeel <model-file> [options]

  --port <n>       Server port (default: 8800)
  --no-open        Don't open browser automatically
  --json           Output model info as JSON (no server)
  -h, --help       Show help
  -v, --version    Show version

How It Works

1. Reads the model file and auto-detects the format
2. Parses the binary structure using a custom built-in parser
3. Builds a graph representation of the model architecture
4. Starts a local HTTP server and opens the viewer in your browser
5. The viewer renders the graph as interactive SVG with pan/zoom

ONNX

Custom protobuf wire-format decoder that extracts the computation graph — nodes, edges, tensor shapes, and operator attributes — without any .proto schema files.

Safetensors

Reads the JSON header to extract tensor names, shapes, and dtypes. Groups tensors by layer prefix to reconstruct the architecture.

GGUF

Binary format parser that reads the GGUF v2/v3 header, metadata key-value pairs, and tensor descriptors including quantization types.

Comparison

| Tool | Deps | Formats | Graph | CLI | |------|------|---------|-------|-----| | mlpeel | 1 optional | 3 | Yes | Yes | | Netron | Electron | 60+ | Yes | No | | onnx-tool | pip | 1 | Partial | Yes |

License

MIT