CodeGemma
GoogleGoogle's code-specialized Gemma family: a 2B pretrained model for fast fill-in-the-middle completion, a 7B pretrained model for infilling and natural language, and a 7B instruction-tuned variant for code chat. All are further-trained on 500 billion to a trillion tokens of public code, mathematics and synthetic data.
Largely superseded for coding by later families.
Openness
3 high confidence- weights
- open
- data
- closed
- code
- closed
- license
- Gemma-Terms-of-Use(custom,use-restricted,non-OSI)
CodeGemma ships under Google's custom Gemma terms: open weights, but no open training data or code. The terms limit commercial use rather than forbid it, which is why the license reads as open weights rather than something more restrictive.
- https://ai.google.dev/gemma/terms recorded 2026-09-29
The Gemma Terms of Use page restricts use: "for the restricted uses set forth in the Gemma Prohibited Use Policy" and lets Google "restrict (remotely or otherwise) usage"; a custom license, not OSI-approved.
- https://huggingface.co/api/models/google/codegemma-7b-it recorded 2026-09-29
The Hugging Face API for google/codegemma-7b-it lists "license:gemma", "license_link":"https://ai.google.dev/gemma/terms" and "gated":"manual", so the weights are downloadable after an access request under the Gemma terms.
- https://huggingface.co/google/codegemma-7b-it recorded 2026-09-29
The model card describes the training corpus only in prose: "further trained on an additional 500 billion tokens of primarily English language data from publicly available code repositories, open source mathematics datasets and synthetically generated code". Searched the body for a dataset link, training code repository and open source license: none; the corpus and training code are not published.
Adoption
2 medium confidenceAdoption is measured as Hugging Face downloads summed across the five CodeGemma repositories declared for this product, most of them for the 2B base model. The two 1.1 variants each account for only a small fraction of that traffic.
- https://huggingface.co/api/models/google/codegemma-1.1-2b recorded 2026-08-12
332 downloads in the trailing 30 days for google/codegemma-1.1-2b
- https://huggingface.co/api/models/google/codegemma-1.1-7b-it recorded 2026-08-12
50 downloads in the trailing 30 days for google/codegemma-1.1-7b-it
- https://huggingface.co/api/models/google/codegemma-2b recorded 2026-08-12
19,078 downloads in the trailing 30 days for google/codegemma-2b
- https://huggingface.co/api/models/google/codegemma-7b recorded 2026-08-12
4,262 downloads in the trailing 30 days for google/codegemma-7b
- https://huggingface.co/api/models/google/codegemma-7b-it recorded 2026-08-12
4,573 downloads in the trailing 30 days for google/codegemma-7b-it
Capability
3 medium confidenceA 2024 code-completion family that is competent for its size but well below current coding models. The Hugging Face release post's comparison table puts CodeGemma 7B at 40.13 on HumanEval Python, behind DeepSeek Coder 7B at 45.83 and StarCoder 2 15B at 44.15, despite 500B tokens of extra training.
- https://github.com/huggingface/blog/blob/main/codegemma.md recorded 2026-08-13
"CodeGemma-7B outperforms similarly-sized 7B models except DeepSeek-Coder-7B on HumanEval"; comparison table - CodeGemma 7B Python 40.13 / JavaScript 43.06, DeepSeek Coder 7B 45.83 / 45.9, StarCoder 2 15B 44.15 / 44.24, CodeGemma 2B 27.28 / 29.94; 2B base for infilling, 7B base, 7B instruct
- https://huggingface.co/blog/codegemma recorded 2026-09-29
The Hugging Face release post states "CodeGemma-7B outperforms similarly-sized 7B models except DeepSeek-Coder-7B on HumanEval" and its table lists CodeGemma 7B at "40.13" against DeepSeek Coder 7B at "45.83" and StarCoder 2 15B at "44.15", with CodeGemma 2B at "27.28".
Verified 2026-09-27