Google DeepMind and Kaggle opened the Gemma 4 Developer Agent Competition, asking teams to turn the open Gemma 4 model into a software-engineering agent that can run offline on consumer hardware. The published prize pool is $100,000.
The brief is explicit. Today's strongest coding agents lean on large cloud APIs. The contest asks whether the same loop — navigate a repo, draft a fix, pass tests — can fit in a quantized model that lives on a normal machine.
What the agent must do
Entrants start from gemma-4-31b-it-qat-w4a16-ct and may add LoRAs, prompts, sub-agents and skills in Google's Agent Development Kit format. The delivery is a submission.zip with agent.yaml at the root.
Scoring follows SWE-Bench logic: apply the patch, run official tests, PASS or FAIL. Agents get a 12-hour budget for the full task set, including sandbox setup.
Dates and prizes
The competition started on September 23, 2026. Rule acceptance and team mergers close November 25. Final submissions are due December 2. Research papers must land by November 12.
- 1st: $37,000
- 2nd: $18,000
- 3rd: $10,000
- Paper track: $35,000, with a spotlight at a Google event tied to NeurIPS
The $100,000 figure is the official sum of both tracks on the Kaggle prize page. A higher “over $110k” line on the @googlegemma account does not match that page.
Why it matters
The editorial point is not the cash. It is the claim that a capable coding agent no longer has to live behind an expensive API. If a quantized Gemma 4 can close real SWE-Bench-style bugs, the ceiling for what fits on a developer's desk moves again.
By the morning of September 29 the Kaggle page already listed thousands of entrants and hundreds of active teams. That measures interest, not quality.
Sources
Image credit: Google Gemma / X
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked against public posts on X and sources available on the internet. Consult the original sources for full context.
Por GeekikiBot
