Kaggle

Ready to build with Google DeepMind's Gemma 4? Following today's Gemma 4 Good hackathon kickoff, we’ve added the Gemma 4 26B and 31B models to Kaggle Benchmarks! Experiment with their multimodal capabilities by handling text and images. See how they perform on your custom evaluation sets. Create your own benchmark and start evaluating the models now: https://lnkd.in/eMFrztr2

Now available on Kaggle: Gemma 4 🤖 In partnership with Google DeepMind, we’re launching the Gemma 4 Good hackathon. In this competition, you’ll leverage local frontier intelligence, native function calling, and multimodal understanding to tackle world-pressing issues. Whether you are optimizing edge-based models for Health & Sciences and Global Resilience, or reimagining the Future of Education and Digital Equity, your solution should demonstrate how Gemma 4 can create meaningful change. Hackathon Details: - Total Prize Pool: $200,000 - Submission Deadline: May 18, 2026 Your work will help show the world that when the right tools are accessible to everyone, the possibilities for positive change are truly endless. 👉 Join the hackathon: https://lnkd.in/eeqqGME9  👉 Gemma 4 on Kaggle: https://lnkd.in/ejrYb3_k

We’ve shipped several important updates to Kaggle Benchmarks. 🤖 Launch of OpenAI models & improved model management: Kaggle Benchmarks now supports the OpenAI ecosystem. You can now evaluate models like GPT-5.4 and GPT-OSS directly alongside other state-of-the-art models. To help you keep pace with the frontier, we’ve also streamlined how you find, run, and track models: - Model selection panel: Easily find and select the specific models you want to test on your task page. - Cross-user runs: You can now run tasks created by other users on models that haven’t yet been evaluated on. - Automated notifications: "Follow" any benchmark to get an alert the second a new model or version hits the leaderboard. 🛠️ Rich token usage, cost, and latency metrics to analyze: A model’s accuracy is only half the story. To help you analyze performance deeper, we’ve added comprehensive metadata to the SDK (see Github repo- https://lnkd.in/eJH5pWaK): - Track token usage: Get exact counts for both “input_tokens” and “output_tokens” on every prompt - Monitor costs: Instantly view “input_tokens_cost” and “output_tokens_cost” to keep your evaluation budgets in check - Measure speed: See the “total_backend_latency” for real-world performance metrics 🏗️ Streamlined benchmark creation and management process:  We’ve unified the creation and editing interface on the benchmark page. - New task panel: A unified interface for managing tasks and running individual models. - Default import of models: When you add a task to a benchmark, you can automatically add all the models a task was previously evaluated on. - Task versioning: You can now see the task versions on your benchmarks and easily update them. Ready to see how the latest models stack up against your custom tasks? Explore Kaggle Benchmarks: https://lnkd.in/eMFrztr2

We are going live in an hour! ⏰ Join us here:  https://lnkd.in/gptwESkz 

The pace of building and deploying AI agents is faster than ever. Now there's a way to evaluate them that keeps pace. Kaggle is launching an experimental MVP Standardized Agent Exams (SAE) — a lightweight, zero-setup way for your AI agent to take a standardized exam and get a score published on a leaderboard instantly. Most benchmarks today are built for foundation models or require fixed harnesses. SAE is built for developers who are deploying agents right now, and extends our work across Game Arena and Benchmarks to advance rigorous, trustworthy evaluation for the GenAI era. What's on the SAE? The first component of the 16-question exam covers the two dimensions that matter most for real-world agent deployment: Reasoning: tests whether your agent can think through multi-step problems Adversarial safety: evaluates whether it responsibly handles tricky or manipulative prompts A zero-setup exam your agent takes on its own: Your agent registers with a single API call - just a name and description, no Kaggle account needed. After the exam, it instantly receives a score, a public report card, and its rank on a live leaderboard. Try it and tell us what you think. 👉 Learn more: https://lnkd.in/eUVAEZUW

Can we truly benchmark AGI? 🧠 Two weeks into the Measuring Progress Toward AGI - Cognitive Abilities hackathon, the benchmarks being built by the Kaggle community are already incredible. To help refine your submissions and ensure they align with the core research goals, we’re hosting a live deep-dive session and AMA on the Kaggle YouTube channel (https://lnkd.in/gv6sVRe5). What we’re covering: - 20-Min Deep Dive into the paper and what we’re looking for in the hackathon - 20-Min Live AMA: Your chance to ask the team anything about the hackathon or the paper The Panel: Nicholas Kang (Kaggle Product Manager), Oran Kelly (Product Manager, Google DeepMind) and Ryan Burnell (Staff Research Scientist, Google DeepMind and co-author, Cognitive Framework paper) Set a reminder for the livestream here: https://lnkd.in/gptwESkz Whether you’re climbing the leaderboard or just interested in the future of AGI evaluation, we’d love to see you there. 🚀 

From disease outbreak detection in West Africa to on-device TB screening - the MedGemma Impact Challenge winners show what's possible when developers and clinicians build together with open-weight models. Congrats to all the winners and thank you to all participants. 👏👇

Real intelligence isn’t about memorizing answers - it’s knowing what to do when the problem changes. Today’s AI systems excel at what they were trained to do, but often fall short when faced with something unfamiliar. Most benchmarks reward pattern recognition, not genuine problem-solving. ARC Prize 2026, in partnership with ARC Prize Foundation, challenges you to build adaptive AI through three connected competitions in the ARC environment. Develop approaches that learn quickly, generalize well, and solve problems never seen before. - ARC-AGI-2:  Predict outputs for novel reasoning tasks your system has never encountered. - ARC-AGI-3: Tackle a harder interactive benchmark requiring exploration and multi-step reasoning, with H100 GPUs and milestone checkpoints in June and September. - Paper Track: Contribute qualitative insights and novel approaches that advance our understanding of generalization. 💰 $2M Prize Pool ⏰ Entry Deadline: October 26, 2026  Compete in one or all three ARC Prize 2026 competitions to help move AI closer to systems that learn like people do: flexible, efficient, and ready for new challenges. ARC-AGI-2: https://lnkd.in/eP4pt6qw ARC-AGI-3: https://lnkd.in/emEDxK4V  Paper Track: https://lnkd.in/ekBe2p9h

We're proud to support innovative integrations! adaption's new data remastering tool lets you take any Kaggle Dataset, enhance it, and share the remastered version back with the community. Learn more 👇

AI is moving fast, and the way we build with it is changing. It’s no longer just about the leaderboard - it’s about the creative ways we apply LLMs to real-world problems. That’s why we’re excited to introduce Community Hackathons on Kaggle - a free, self-serve way for you to host your own AI challenges. Whether you're an educator, a meetup lead, or just have a big idea, you can now have the same tech the pros use to build, judge and award prizes (up to $10k!). We’ve built in the tools interactive notebooks, data hosting, and a new project gallery so you can focus on the innovation. Check out the full details: https://lnkd.in/e_UUkaWz