Google Launches Gemma 3: Top Single-GPU Multimodal Model for Mobile, 27B Beats o3-mini

Google Gemma 3: Open-source, multimodal AI with 128k context, 27B model outperforms o3-mini. Runs on single GPU/TPU, supports 140+ languages, and excels in math, coding, and vision tasks.

Mar 12, 2025

∙ Paid

"AI Disruption" Publication 5000 Subscriptions 20% Discount Offer Link.

Gemma 3: Google's new open model based on Gemini 2.0

Google Gemma 3 Family is Here!

Just now, at the Paris Developer Day, the open-source Gemma series models officially iterated to the third generation, natively supporting multimodal capabilities and a 128k context length.

This time, Gemma 3 has open-sourced four parameter sizes: 1B, 4B, 12B, and 27B. The most crucial point is that the models can run on a single GPU/TPU.

In the LMArena competition, Gemma 3 scored a high 1339 ELO, defeating o1-preview, o3-mini high, and DeepSeek V3 with just 27B parameters, making it the best open-source model second only to DeepSeek R1.

AI Disruption

Google Launches Gemma 3: Top Single-GPU Multimodal Model for Mobile, 27B Beats o3-mini

Google Gemma 3: Open-source, multimodal AI with 128k context, 27B model outperforms o3-mini. Runs on single GPU/TPU, supports 140+ languages, and excels in math, coding, and vision tasks.

This post is for paid subscribers