gemma-3-27b-it

Gemma 3 27B IT is Google DeepMind's largest instruction-tuned model in the Gemma 3 family, released in March 2025. Built on the same research foundation as Gemini, it is a multimodal vision-language model that accepts text and image inputs (896x896 resolution, encoded to 256 tokens per image) and generates text with a 128K-token context window and up to 8K output tokens. The model uses a decoder-only Transformer with Grouped-Query Attention, alternating 5:1 local/global attention for efficient long-context inference, and integrates a 417M-parameter SigLIP vision encoder.

The 27B variant was pre-trained on 14 trillion tokens spanning web text, code, mathematics, and images across 140+ languages, then refined through knowledge distillation and reinforcement learning for instruction following, reasoning, and tool use. On LMSYS Chatbot Arena it achieved 1338 Elo, ranking among the top open-weight models and outperforming much larger peers such as DeepSeek-V3 and Qwen2.5-70B. Benchmark results place it near Gemini 1.5 Pro on tasks including MMLU-Pro (67.5), MATH (89.0), HumanEval (87.8), and MMMU (64.9).

Gemma 3 27B IT is released under the Gemma Terms of Use with open weights on Hugging Face, Kaggle, and Vertex AI Model Garden. It supports deployment via Transformers, llama.cpp, and MLX, making it suitable for chatbots, document understanding, OCR, code generation, and multilingual applications on both cloud and consumer hardware when quantized.