Improving Hugging Face Model Access for Kaggle Users

Hugging Face and Kaggle are tightening the cord between their two ecosystems, and the update is less flashy than useful. Starting today, you can go from a Hugging Face model page — say Qwen/Qwen3-1.7B — straight into a Kaggle notebook with the loading code already written for you. Click "Use this model," pick Kaggle, and you're in an environment ready to run. The reverse works too: Kaggle model pages now carry a "Code" button that opens the same shortcut in the other direction.

What's more interesting is the automation happening quietly in the background. If someone runs a Kaggle notebook that pulls a model from Hugging Face Hub, and that model doesn't yet have a page on Hugging Face, one gets created automatically. No extra steps, no code changes required. Make the notebook public and it shows up under the "Code" tab of the corresponding Kaggle model page, which effectively turns every public Kaggle notebook into a piece of documentation and inspiration for that model. Over time, as usage grows, this should build a much richer library of real-world examples sitting right next to the models themselves.

Private and gated models get handled sensibly rather than glossed over. Private models still require your HF_TOKEN through Kaggle's secrets menu, same as before, and won't trigger an auto-generated public page — which makes sense, since nobody wants their internal fine-tune showing up on a public model card by accident. Consent-gated models still route through Hugging Face's usual access-request flow before anything loads.

The part still in progress is arguably the trickiest: getting this to work inside Kaggle competitions that require offline notebook submissions. Kaggle is treating that carefully, and for good reason — competition integrity depends on avoiding data leakage and contamination, and letting models phone home to Hugging Face mid-competition could undermine exactly the kind of rigor Kaggle prides itself on. The teams say a proper solution is still a few months out. Given how much weight Kaggle places on competitions as a benchmark for evaluating generative AI, that caution tracks.