Google DeepMind Publishes Gemma 3 Technical Report Detailing 27B Open Model

Google DeepMind has released the technical report for Gemma 3, its latest open model that features 27 billion parameters with support for vision, text generation, and over 140 languages. The report details the model's architecture, training methodology, and benchmark performance, which is competitive with significantly larger open-source models.

Architecture

Gemma 3 uses a dense transformer architecture with 27 billion parameters — a deliberate choice to remain at a size that can run on consumer-grade GPUs with 24GB of VRAM. The model supports both text-only and vision-language tasks through a unified architecture that processes images and text within the same model, rather than using separate vision encoders. This approach simplifies deployment and reduces the overhead of running multimodal AI applications.

Performance

On standard benchmarks, Gemma 3 performs competitively with models that are 2-3x larger, including Meta's Llama 3.3 70B and Alibaba's Qwen 3.5 72B. The model achieves particularly strong results on multilingual benchmarks, reflecting Google's investment in training data diversity across its 140+ supported languages. On coding benchmarks, Gemma 3 scores within a few percentage points of the best open-source models, making it a viable option for code-assisted development workflows.

Open Model Ecosystem

Gemma 3 is available under Google's permissive Gemma license, which allows commercial use, fine-tuning, and redistribution. The release continues Google's strategy of maintaining a competitive position in the open-source AI ecosystem alongside Meta's Llama, Alibaba's Qwen, and Mistral's models. For developers, the practical significance is that a 27B model that runs on a single consumer GPU can now perform tasks that required much larger models — or expensive API calls — just months ago, further reducing the cost of building AI-powered applications.