Benchmarks, Game Arenas, Writeups, and More: Some Exciting New Updates from Kaggle
Some cool new things are now live on Kaggle.
While we all know Kaggle as the go-to platform for machine learning competitions, the team has recently rolled out some exciting initiatives that go well beyond traditional contests. There have been so many lately that it was getting hard to keep track, so I’ve put together a simple list of the features I’m genuinely excited about and shared them all here in one place.
1. Putting AI to the Test with Kaggle Benchmarks
Let’s start with what I believe is the most transformative update: a new system for AI evaluation. This is a massive step forward for the entire AI community.
Benchmarks, while important, are difficult to create and maintain, especially for smaller teams or solo researchers. This is because building a reliable benchmark requires the correct combination of the right datasets, evaluation scripts and not to forget, the compute. That’s exactly where Kaggle Benchmarks come in. It brings more than 70 verified leaderboards from top research benchmarks like MMLU, SciCode, MathVista, and Meta’s Multiloko directly to the community.
The fact that these are fully reproduced by Kaggle ensures transparency and trust. Even better, you can create and run custom evaluations on popular LLMs without worrying about infrastructure or compute costs since Kaggle handles all of that for you.
A great example is Meta’s Multiloko, which tests multilingual capabilities across 31 languages and uses a private, held-out test set to prevent data leakage.
Looking ahead, Kaggle Benchmarks will support multimodal tasks, community-built benchmarks, and even stress-testing competitions for models. All of this is a great step in advancing trustworthy AI evaluation.
2. Let the Games Begin: AI vs. AI in the New Game Arena
Building on the foundation of Benchmarks, the Kaggle Game Arena takes AI evaluation to the next level by allowing models and agents to compete head-to-head in strategic games.
If you’ve ever tried Kaggle Simulations, the Game Arena will feel familiar, but with a key difference. Instead of community teams competing, this is about ranking top AI models like Gemini 2.5 Pro, Claude Opus 4, o3, Grok 4 and others on live, ongoing benchmarks that anyone can follow.
The arena kicked off with a chess tournament, where models compete in streamed, replayable matches with results tracked on a dynamic leaderboard. Games provide a better testbed for models because they involve behaviors like strategic planning, reasoning, and adaptation, offering a truer measure of an AI’s capabilities.
Kaggle has hinted at even more games in the future, including multiplayer titles, video games, and real‑world simulations. If the first event is any sign, Game Arena is only going to get more exciting from here.
3. Hackathons Are Here (And They’re Not Just About Predictions)
Kaggle Hackathons are a new type of competition that goes beyond traditional predictive modeling. They might ask you to build an app, create a new metric, creatively use an LLM, or even produce an educational YouTube video. This format allows for qualitative evaluation and highlights a wider range of developer skills.
A nice bonus is that you can submit projects even if they’re built off‑platform. Entries are submitted as Hackathon Writeups consisting of rich, multimedia posts where you can link to and describe your project in detail. Beyond participating, you can also host your own Kaggle Hackathon. While these events don’t award points or medals, they do come with prizes and genuinely interesting challenges. Recent examples include the Meta Kaggle Hackathon and the Gemma 3n Impact Challenge.
4. Show Off Your Work with the New Kaggle Writeups
I talked about the Writeups above as a way to submit Hackathon entries, but you’re not limited to that. You can use them for competition solutions, personal projects, or even professional work. Basically, it’s a cleaner, long‑form format to tell the story behind your data science work.
For instance, here are two write-ups that I have recently submitted to two hackathons. They look nicer than forum posts, support videos and images, and let you link to external resources like GitHub or Hugging Face with rich embedded cards. A lot of times, it becomes hard to find a competition solution and one has to dig through old forum threads. Writeups make that so much easier.
5. Need More Power on Kaggle? Link Colab for a GPU Boost
Kaggle and Colab have long helped the GPU‑poor by providing free access to powerful hardware. But for anyone who constantly needs more GPU power, Kaggle now has an experimental feature that can help. You can now link an active Colab Pro or Pro+ subscription to boost your weekly GPU quota on Kaggle. Also, this bonus doesn’t affect your Colab compute units and runs on Kaggle’s reliable hardware. Linking is simple: just open the Kaggle notebook editor, choose Link to Colab from the File menu, and follow the prompts for an instant upgrade.
You can read more about it here.
6. Your Workflow Just Got a Lot Smoother with GitHub & Colab Sync
Kaggle has made working with notebooks across platforms a lot smoother. For instance, Colab notebooks now sync smoothly with Kaggle, and you’ll even see alerts when there are updates so you can pull changes without hassle. You can also bring notebooks straight from GitHub, a URL, or even a local file. I like that you can either run it immediately or just save it for later. It’s especially handy when you’ve been working on something locally and want to move it to Kaggle without rerunning the whole thing first.
7. Hugging Face Models Now Play Nice with Kaggle
Hugging Face and Kaggle finally teamed up in a way the community has been asking for. Now, Hugging Face models are easier to find and use right inside Kaggle. From any model page on Hugging Face, you can click Use this model and send it straight to a Kaggle notebook with the code already set up. Even better, if you use a Hugging Face model in one of your Kaggle notebooks, Kaggle will automatically create a model page for it. Once your notebook is public, it shows up under that model’s Code tab, turning it into a growing library of real community examples.
What these updates mean for the community
While there are a few other updates from the Kaggle team, these are the ones that really stuck out to me. Each one meaningfully improves the experience on the platform and offers a more comprehensive and powerful ecosystem for everyone in the data science community.
