Breaking

Google Backs One-Day Gemma 4 Hackathons on Kaggle

June 24, 2026 at 15:06 EDT

Open Source
Foundation Models
AI Agents

Google DeepMind's Gemma team announced on June 24, 2026, that it will sponsor community-led, one-day hackathons on Kaggle where developers build with the open-weight Gemma 4 model. Prospective organizers can apply, and participants use Gemma 4 to create lightweight tools and tackle challenges specific to their communities.

June 24, 2026 · Google DeepMind — Gemma

Developer-Led Hackathons Built on Gemma 4

A new program sponsors single-day, community-run hackathons on Kaggle, inviting developers to launch their own events and build lightweight, edge-ready tools with the open-weight Gemma 4 model.

$200K

Total prize pool — Gemma 4 Good Hackathon (spring 2026)

~$10K

Prize cap for sponsored community hackathons

140+

Languages covered by Gemma 4 models

AIME 2026 Math — Gemma 4 vs. Prior Generation

Score in thinking mode — the headline leap of the new generation

20.8%

Gemma 3
27B

89.2%

Gemma 4
31B

88.3%

Gemma 4
26B A4B

A 4× jump. Gains span benchmarks: MMMLU 67.6% → 85.2%, LiveCodeBench 29.1% → 80.0%, Arena AI 1365 → 1452.

One Family, Four Configurations

E2B / E4B

Edge / Mobile

Per-Layer Embeddings, 128K context, low-memory operation

12B

Unified Multimodal

Encoder-free across text, image, audio

26B A4B

Mixture-of-Experts

~4B active parameters

31B

Dense

High performance

What developers value

Offline, edge-ready, privacy-focused builds
Fit with real-world problem themes
Easy result sharing on Kaggle
QAT models cut memory — E2B from 11.4GB to 2.9GB at 4-bit

Open concerns

Memory burden of the 26B / 31B class
Early fine-tuning difficulty
Need for attention to data quality and safety

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.