stprobe

stprobe is ffprobe for .safetensors.

It is a small Rust CLI for answering:

"What is inside this safetensors file?"

Point it at a model file and it prints a stable plain-text summary without loading tensor payloads into memory.

It works with local files and http(s) URLs, including Hugging Face resolve links.

Why stprobe:

• fast: header-only inspection, including huge remote files over HTTP range requests
• small: prebuilt release archives are about 1.4 to 1.6 MB across supported platforms
• easy: one command, no Python environment, stable plain-text output

Because it only probes the safetensors header and metadata, it is fast even for very large remote files.

stprobe shows:

• file path
• file size
• tensor count
• metadata
• per-tensor name, dtype, shape, parameter count, and byte size
• total parameters
• total tensor bytes
• dtype breakdown

It is built for model inspection, debugging, issue reports, and quick shell use.

• no Python environment
• no notebook setup
• no ad hoc parsing script
• header-only inspection with the official safetensors crate
• remote probing over HTTP range requests instead of downloading the full file
• stable plain-text output that is easy to read, diff, grep, and paste

Install

Use a prebuilt binary from GitHub Releases if you just want to run stprobe without installing Rust:

• Download the archive for Linux x86_64, macOS Intel, macOS Apple Silicon, or Windows x86_64 from: https://github.com/Xyc2016/stprobe/releases/latest
• Extract it and run stprobe --version

Linux x86_64 example:

curl -L https://github.com/Xyc2016/stprobe/releases/latest/download/stprobe-x86_64-unknown-linux-gnu.tar.gz | tar -xz
./stprobe --version

Or install from crates.io:

cargo install stprobe

Quick Start

stprobe all-MiniLM-L6-v2.safetensors

Example:

$ stprobe all-MiniLM-L6-v2.safetensors
File: all-MiniLM-L6-v2.safetensors
Size: 90868376 bytes
Tensors: 104
Parameters: 22713728
Tensor-Bytes: 90856960

Metadata:
  format = pt

DType Breakdown:
  F32: 90852864 bytes
  I64: 4096 bytes

Tensors:
  embeddings.position_ids
    dtype: I64
    shape: [1, 512]
    numel: 512
    bytes: 4096

  embeddings.word_embeddings.weight
    dtype: F32
    shape: [30522, 384]
    numel: 11720448
    bytes: 46881792

Fast On Huge Remote Files

You can also probe a large Hugging Face model directly over HTTP without downloading the full .safetensors file first. This example pins a specific Hugging Face revision so the output stays reproducible:

time stprobe https://huggingface.co/Comfy-Org/flux1-dev/resolve/0f6b956e6e2e041fb73d079b72ec0e761506f601/flux1-dev-fp8.safetensors

Example:

$ time stprobe https://huggingface.co/Comfy-Org/flux1-dev/resolve/0f6b956e6e2e041fb73d079b72ec0e761506f601/flux1-dev-fp8.safetensors
File: https://huggingface.co/Comfy-Org/flux1-dev/resolve/0f6b956e6e2e041fb73d079b72ec0e761506f601/flux1-dev-fp8.safetensors
Size: 17246524772 bytes
Tensors: 1442
Parameters: 16871188965
Tensor-Bytes: 17246298324

Metadata:
  modelspec.architecture = Flux.1-dev
  modelspec.author = Black Forest Labs
  modelspec.date = 2024-08-01
  modelspec.description = A guidance distilled rectified flow model.
  modelspec.hash_sha256 = 0x2f3c5caac0469f474439cf84eb09f900bd8e5900f4ad9404c4e05cec12314df6
  modelspec.implementation = https://github.com/black-forest-labs/flux
  modelspec.license = FLUX.1 [dev] Non-Commercial License
  modelspec.sai_model_spec = 1.0.1
  modelspec.thumbnail = data:image/jpeg;base64,/9j/4AAQSkZJRgABAQAAAQABAAD/...
  modelspec.title = Flux.1-dev

DType Breakdown:
  F16: 247300608 bytes
  F32: 335278740 bytes
  F8_E4M3: 16663718976 bytes

Tensors:
  text_encoders.clip_l.logit_scale
    dtype: F32
    shape: []
    numel: 1
    bytes: 4

  ...

  text_encoders.t5xxl.transformer.shared.weight
    dtype: F8_E4M3
    shape: [32128, 4096]
    numel: 131596288
    bytes: 131596288

stprobe https://huggingface.co/Comfy-Org/flux1-dev/resolve/0f6b956e6e2e041fb73d079b72ec0e761506f601/flux1-dev-fp8.safetensors  0.01s user 0.02s system 2% cpu 1.277 total

The modelspec.thumbnail value is abbreviated here for readability.

Example timing from one local machine using the optimized build. Exact results vary with network, filesystem cache, shell, and hardware.

Why Not A One-Off Script

You can inspect a safetensors file with a short Rust or Python script. stprobe is better when you want something you can keep using:

• one command instead of re-editing one-off scripts
• stable output for bug reports, CI logs, and terminal use
• no Python packaging, no notebook, no local helper code
• uses the official safetensors crate instead of custom header parsing
• reads only the safetensors header and metadata when that is all you need

Build From Source

cargo install --path .

Or run it without installing:

cargo run -- model.safetensors