Gemma 2 9B

Gemma 2 9B is the 9-billion-parameter member of Google DeepMind's Gemma 2 family, released June 27, 2024 in both base and instruction-tuned variants. Trained on 8 trillion tokens with knowledge distillation from larger models, it outperformed Llama 3 8B and other similarly sized open models on launch benchmarks (75.2 MMLU, 51.8 HumanEval for the 9B IT variant), with an 8K context window. It distills research from the Gemini program into a package that fits on a single high-end GPU.

Google distributes the weights on Hugging Face and Kaggle under the Gemma license (commercial use permitted with some restrictions), plus API access via AI Studio and Vertex AI. The honest picture: Gemma 2 is now the prior generation — Gemma 3 superseded it — and the 8K context window looks small next to current long-context models. The Gemma license is not fully permissive open source, so check its terms for redistribution. Still, for teams that standardized on Gemma 2, the 9B remains the sweet spot of capability per VRAM.