Gemma 3

Description: Google's lightweight open-source models based on Gemini 2.0 technology

Website: https://ai.google.dev/gemma

Google released Gemma 3 on March 12, 2025 as a family of lightweight open-source models. Built on the same technology as Gemini 2.0, optimized for single-GPU operation.

Model Sizes

• 270M and 1B: Text only, ideal for edge devices
• 4B, 12B, 27B: Multimodal (text + images), state-of-the-art for their size

Performance

Sometimes surpasses larger models like Llama3-405B, DeepSeek-V3 and o3-mini in evaluations. Trained with 2T to 14T tokens on Google TPUs.

Features

• Context: 128K tokens (32K for smaller models)
• Languages: Over 140 languages
• Function calling for complex tasks
• Multimodal: Text and images (larger variants)

Availability

Over 100 million downloads, 60,000+ community variants. Available on Kaggle and Hugging Face. Runs smoothly locally with Ollama, LM Studio or llama.cpp.

Use Case

Perfect for developers who want to run performant models on consumer hardware.