gemma-2-2b-it

Gemma 2 2B Instruct (gemma-2-2b-it) is Google's lightweight instruction-tuned language model from the Gemma 2 family, released in July 2024. Built with the same research foundations as Gemini, it is a decoder-only Transformer with grouped-query attention (GQA), interleaved local (4,096-token) and global (8,192-token) attention layers, and logit soft-capping. The 2B variant was trained on 2 trillion tokens via knowledge distillation from larger teacher models, then post-trained with supervised fine-tuning and reinforcement learning for chat and instruction following.

Despite its small size (~2.6B total parameters), Gemma 2 2B IT delivers strong performance for its class. On LMSYS Chatbot Arena it scores 1,126 Elo, outperforming GPT-3.5-Turbo-0613 (1,116). It achieves 56.1% on MMLU and 36.6% on MBPP after instruction tuning, with competitive results on reasoning, coding, and safety benchmarks. Open weights are available under the Gemma license on Hugging Face, Kaggle, and Vertex AI Model Garden.

Designed for efficient deployment, Gemma 2 2B IT targets on-device and resource-constrained use cases including laptops, edge hardware (NVIDIA Jetson), and cloud inference via TensorRT-LLM and NVIDIA NIM. It supports an 8,192-token context window and uses a simple chat template compatible with Hugging Face transformers.