Hugging Face’s science team got tired of experiment trackers that cost money, lock up data, or require complex setups. Their answer is Trackio, a new open-source Python library that they’re already using internally. It’s designed as a drop-in replacement for Weights & Biases — you can literally change your import statement from import wandb to import trackio as wandb and your existing logging code keeps working.
The library logs metrics locally and visualizes them through a Gradio dashboard you can launch from your terminal with trackio show. What makes this genuinely useful, not just another logging tool, is the sharing model. By passing a space_id to trackio.init(), your dashboard syncs directly to a Hugging Face Space. That means you can embed live training plots into a blog post or documentation using a simple iframe — no accounts, no dashboards, no friction for the person viewing your results.
There’s a practical environmental angle here that matters more than most people admit. Trackio pulls GPU energy usage straight from nvidia-smi, making it trivially easy to quantify power consumption during training and add that data to model cards. The team is pushing for transparency around the energy costs of model training, and baking this into the default tooling is a smart way to make it standard practice rather than an afterthought.
One detail that signals this was built by people who actually train models: they’ve thought about throughput. The library lets you control when tensors move from GPU to CPU during logging, so you can track model states mid-training without tanking performance. It’s a small thing that only matters if you’ve burned hours debugging slow training loops. Trackio is on PyPI now, installable with a single pip command, and the Spaces integration means private dashboards for your org are free.
