Google Gemma 4 Developer Agent Competition 2026 on Kaggle
Kaggle competition from Google DeepMind to train Gemma 4 into an offline coding agent. $100,000 in prizes; join by 25 November 2026.
The Google Gemma 4 Developer Agent Competition is an online Kaggle competition hosted by Google DeepMind. Participants post train the open Gemma 4 model into an autonomous coding agent that can move through large codebases and draft fixes for real software issues. The goal is an agent that works offline on everyday hardware instead of relying on large cloud models.
The total prize money is $100,000. The main leaderboard competition pays $65,000: $37,000 for first place, $18,000 for second and $10,000 for third. An optional paper track pays $35,000 for original research on the methods behind the agents, and top papers will be shown at a Google event such as a NeurIPS expo workshop. If a team wins, the money is split evenly between its members unless the team agrees on another split.
The competition is open worldwide to people aged 18 or over, or the age of majority where they live. Residents of Crimea, the so called Donetsk and Luhansk People's Republics, Cuba, Iran and North Korea cannot enter, and neither can people under US export controls or sanctions. You can take part alone or in a team, using one Kaggle account each.
To enter, sign in to Kaggle, join the competition and accept the rules. Submissions are zip files with an agent config built on the Google ADK Agent Config format. Agents can only use the tools the competition harness provides, such as running shell commands, reading and editing files and searching a code graph. Patches are scored like SWE-Bench: a fix counts only if the issue's tests pass.
The competition started on 23 September 2026. Research papers are due on 12 November 2026. The entry deadline and the team merger deadline are both 25 November 2026, and final submissions close on 2 December 2026. All deadlines are at 11:59 PM UTC.
Quick answers
What do I actually submit?
A zip archive with an agent config: system prompts, custom tools, skills and optional LoRA adapters, with an agent.yaml file at the root. The agent must use one of four listed Gemma 4 model variants.
How is my agent scored?
Like SWE-Bench. Each patch your agent writes is applied to a real repository and that issue's tests are run. Your score is the share of issues that pass, and your agent has 12 hours for all tasks.
Can I win money without topping the leaderboard?
Yes. A separate paper track pays $35,000 for research write ups on topics such as fine tuning, code comprehension and graph reasoning. Papers are due on 12 November 2026.