Google: Gemma 2 9B

Gemma 2 9B is a 9-billion-parameter open-weights language model from Google, released on June 27, 2024 alongside a 27B sibling in Google's Gemma 2 announcement. The model card describes Gemma as a family of lightweight open models built from the same research as Gemini: text-to-text, decoder-only language models for English, with open weights. This is the pretrained checkpoint; an instruction-tuned gemma-2-9b-it also exists. Google's Gemma page now lists four generations, Gemma 1 to Gemma 4, so Gemma 2 is a previous generation but remains downloadable.

Key Features

• Size and data: 9B parameters, trained on 8 trillion tokens (the 27B model used 13 trillion), according to the card.
• Open weights: both pretrained and instruction-tuned variants have open weights. The Hugging Face repository is gated: you must log in and accept Google's Gemma license.
• Intended uses: text generation, chatbots, summarization, NLP research and language-learning tools, including on laptops and desktops.
• Vendor benchmarks: the card reports, for the pretrained 9B model, MMLU 71.3 (5-shot), HellaSwag 81.9, GSM8K 68.6, HumanEval 40.2 pass@1 and MATH 36.6. These are Google's figures. Google's announcement also said the 9B model outperforms Llama 3 8B and other open models of its size.

Use Cases

• Running a capable small model locally or on a single GPU.
• Fine-tuning a base model for domain text generation or summarization.
• Research baselines for small English language models.

Limitation: the card does not state a context length, and Gemma 2 is English-oriented, so check your own multilingual needs.

Pricing and availability

The weights are free to download under the Gemma license terms. Google's announcement also pointed to Google AI Studio for hosted access; this entry quotes no hosted price because none was checked on a first-party page.

Getting Started

1. Accept the Gemma license on the Hugging Face model page.
2. Load google/gemma-2-9b with Transformers, or use the instruction-tuned google/gemma-2-9b-it for chat.
3. For current work, compare against the latest Gemma family on Google's Gemma page.

FAQ

Is Gemma 2 9B open source?

It has open weights under Google's Gemma license, which requires accepting terms. Read the license for your use case.

What replaced Gemma 2 9B?

Google's Gemma page lists Gemma 4 as the latest generation, with sizes E2B, E4B, 12B, 31B and 26B A4B. The catalog has Gemma 4 26B A4B.

How many tokens was it trained on?

8 trillion tokens for the 9B model, per the model card.

Is there a larger Gemma 2?

Yes, a 27B model; see Gemma 2 27B.

Alternatives

• Gemma 2 27B: the larger Gemma 2 sibling.
• Gemma 4 26B A4B: a current-generation Gemma model.
• EmbeddingGemma: Google's Gemma-based embedding model.

Conclusion

Gemma 2 9B is a well-documented, openly available small model that is now one generation behind. Use it for existing pipelines or research baselines, and evaluate Gemma 4 for new work. See more in the models category.