Google just dropped Gemma 3, the latest in its open-model lineup for devs.
Google’s AI party is here to flex on your single GPU like it owns the place.
After rolling out Gemma 1 in February and Gemma 2 in May, this new version builds on Google's AI research and packs some serious upgrades.
Gemma models have been blowing up—over 100 million downloads in the past year and 60,000+ variants floating around in what Google calls the “Gemmaverse.”
These models are designed to run fast straight on devices like phones, laptops, and workstations.
Gemma 3 comes in four sizes: 1B, 4B, 12B, and 27B, and it’s built on the same tech that powers Gemini 2.0.
Google is flexing hard on its claim that Gemma 3 is the "world’s best single-accelerator model," meaning it beats out Llama-405B, DeepSeek-V3, and o3-mini in LMArena, all while running on a single GPU or TPU.
- Better text and visual reasoning – It can analyze images, text, and short videos, with 4B+ models supporting a 128k-token context window.
- Multilingual flex – Out of the box, it supports 35+ languages and has pre-trained support for over 140.
- AI workflow automation – Supports function calling and structured output, so devs can build smarter, more agentic AI systems.
- Faster, lighter models – Official quantized versions shrink model size and lower compute costs while keeping accuracy on point.
On the safety side, Google introduced ShieldGemma 2, a 4B image safety checker that flags content across three categories: dangerous, explicit, and violent.
Plus, the company promises “extensive data governance” and alignment with its safety policies through fine-tuning and benchmark evaluations.
In short: Gemma 3 is here to push the limits of on-device AI, and Google wants devs to start experimenting. |
