Format Atlas

Guides

Checks before handing off an AI model

Distinguish weights, computation graphs and inference environments. Deliver dependencies and representative-input results together; successful conversion alone does not establish compatibility or quality.

Which format should you choose?

Choose by the role each file plays in your workflow.

  • Record execution requirements and provenance
  • Assemble the required artifacts
  • Validate inference at the destination

Record execution requirements and provenance

Record source, revision, usage terms, architecture, input/output types and shapes, preprocessing and postprocessing. Match ONNX IR/opset and GGUF tensor types to the target runtime.

Assemble the required artifacts

Preserve ONNX external-weight references. Include required Safetensors shards, indexes, model definitions and other dependencies. GGUF may also need split or auxiliary files. Check whether tokenizers and settings are embedded or supplied separately.

Validate inference at the destination

Use an updated compatible loader to check structure and tensor names, types and shapes. With identical preprocessing and inputs, compare outputs with the source and record tolerances and memory usage. Evaluate conversion and quantization changes separately.

Formats to consider

ONNX

Good for: Exchanging computation graphs with compatible runtimes

Less suitable for: Assuming support for every operator and opset

Safetensors

Good for: Saving and distributing tensor weights

Less suitable for: Assuming weights alone include model definitions and executable code

GGUF

Good for: Delivering inference models to compatible GGML implementations

Less suitable for: Inferring quantization or compatibility from the extension

Before you export

  • Format names or matching hashes do not prove provenance or safety. Do not carelessly execute untrusted code or conversion inputs; check artifact origins.
  • Retain the source model, conversion settings and file inventory. Matching one test input does not establish quality for all inputs or devices.

Related comparisons

Safetensors vs ONNX: weights or model graphsSafetensors stores named tensors. ONNX represents a model with a computation graph, inputs, outputs, and operator versions.GGUF vs Safetensors: inference delivery and weight storageGGUF delivers inference models to compatible GGML-based implementations. Safetensors stores named tensors; model definitions and execution environments are supplied separately.Keras vs ONNX: training saves and inference exports.keras saves model configuration, weights and any optimizer state in Keras. ONNX represents computation graphs for compatible runtimes.

Guides

Checks before handing off scientific arraysTransfer netCDF, HDF5 or Zarr arrays with their meaning, dependencies and reader requirements. Unlike choosing a table format, this requires checking dimensions and missing-value interpretation.

References and specifications