Google: Gemma 2 27B

Gemma 2 27B is the largest model in Google's Gemma 2 open-weights family. The Hugging Face model card describes it as a text-to-text, decoder-only large language model in English, with 27B parameters trained on 13 trillion tokens (the 9B sibling used 8 trillion). Access is gated: you must accept Google's usage terms, and the weights come under the Gemma license. The repo is part of the Gemma 2 release collection, with both pretrained (PT) and instruction-tuned (IT) versions.

Key features

• 27B parameters, 13T tokens: the card lists 13 trillion training tokens for the 27B model.
• Pretrained and instruction-tuned: PT and IT variants exist; this repo is the pretrained model.
• Vendor benchmarks: for the pretrained 27B model (versus 9B), Google reports MMLU 75.2 (71.3), HellaSwag 86.4 (81.9), GSM8K 74.0 (68.6), HumanEval pass@1 51.8 (40.2) and MATH 42.3 (36.6).
• Broad tooling: the card lists Transformers, vLLM, SGLang and Docker Model Runner.
• Ecosystem: the Hub shows 9 adapters, 29 finetunes and 14 quantizations built on this repo at the time we checked.

Use cases

• Text generation, chatbots and conversational AI, summarization and NLP research, as listed on the card.
• Language-learning tools and knowledge exploration.
• A base for your own fine-tuning or quantization experiments.

Pricing and availability

The weights are free to download from Hugging Face after you accept Google's terms; the license is the Gemma license. Running a 27B model needs a capable GPU, and hosted pricing depends on the provider. The card does not state a context length or a deprecation date, and we checked it on 2026-10-29.

Getting started

1. Request access on the Hugging Face repo and accept the terms.
2. Load the model with Transformers, or serve it with vLLM or SGLang.
3. Use the instruction-tuned variant if you need chat behavior; this repo is the pretrained model.
4. Evaluate it on your task against newer open models before adopting.

FAQ

How many parameters does Gemma 2 27B have?

27B, trained on 13 trillion tokens, per the Hugging Face card.

What license does it use?

The Gemma license. Access is gated behind Google's usage terms.

Is this the chat version?

No. This repo is the pretrained model; the instruction-tuned version is a separate variant.

Is it multilingual?

The card describes it as an English-language model.

What is its context length?

The card does not state it.

Alternatives

• Gemma 2 9B: the smaller sibling that needs less hardware.
• Gemma 4 26B A4B: a later-generation Gemma model in this catalog.
• Llama 3.1 70B Instruct: a larger open-weights instruct model.

Conclusion

Gemma 2 27B is a well-documented open-weights base for text generation and fine-tuning, with clear training and benchmark disclosures from Google. Check the model card for terms, compare it with newer Gemma models, and browse the models category.