gemma-2-9b-it

Gemma 2 9B IT is Google's instruction-tuned, open-weight language model from the Gemma 2 family, built on the same research foundation as Gemini. With 9 billion parameters, it targets conversational AI, instruction following, summarization, and reasoning while remaining deployable on consumer hardware. The model uses a decoder-only Transformer architecture with innovations including sliding-window attention, logit soft-capping, and knowledge distillation from larger teacher models.

Released in June 2024, Gemma 2 9B IT delivers strong performance for its size class, outperforming many larger open models on Chatbot Arena (Elo 1187, comparable to GPT-4-0314) and excelling on benchmarks such as MMLU (72.3%), GSM8K (70.7%), and MBPP (62.2%). It supports an 8,192-token context window and runs efficiently in BF16 on GPUs with roughly 18 GB VRAM, or in 4-bit quantized form on 6 GB cards.

The model is released under the Gemma Terms of Use with open weights on Hugging Face and Kaggle. It is English-focused, text-only, and suited for chatbots, content generation, code assistance, and research. Safety evaluations cover toxicity, bias, and memorization, with instruction tuning improving both capability and alignment relative to the base pretrained variant.