Gemma 4

weights

open(Apache-2.0)

data

closed

code

closed

license

Apache-2.0(OSI,but no open data/code)

Gemma 4 ships under Apache-2.0 and is ungated on Hugging Face, unlike Gemma 2 and 3, which remain gated under the Gemma License and its Prohibited Use Policy. The model card describes the pretraining mixture without releasing it and links no training pipeline, so Gemma 4 is an open-weights release rather than a fully open model.

• https://ai.google.dev/gemma/docs/gemma_4_license recorded 2026-09-29

  The Gemma 4 license link redirects to ai.google.dev/gemma/apache_2, canonical "https://ai.google.dev/gemma/apache_2", titled "Apache License 2.0", which carries the standard Apache text rather than a Gemma Terms of Use page. Searched the license body for prohibited use and Gemma Terms: they appear only as site navigation links.

• https://ai.google.dev/gemma/docs/releases recorded 2026-09-29

  Release log listing Gemma 4 "in E2B, E4B, 31B and 26B A4B sizes" on March 31, 2026 and Gemma 4 12B Unified on June 3, 2026. It carries no license statement for the models; the only Apache text is the site footer covering code samples.

• https://arxiv.org/abs/2607.02770 recorded 2026-09-29

  Gemma 4 Technical Report abstract: "We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family." Searched the abstract page for dataset release, training code and open-source: none.

• https://huggingface.co/api/models/google/gemma-4-12B-it recorded 2026-09-29

  Hub record for google/gemma-4-12B-it, the June 2026 12B Unified release, tagged "base_model:finetune:google/gemma-4-12B" and "license:apache-2.0", with gated and private both false.

• https://huggingface.co/api/models/google/gemma-4-26B-A4B-it recorded 2026-09-29

  Hub record for "google/gemma-4-26B-A4B-it" carries the tag "license:apache-2.0" with gated and private both false.

• https://huggingface.co/api/models/google/gemma-4-31B recorded 2026-09-29

  Hub record for the pretrained base google/gemma-4-31B carries the tag "license:apache-2.0" and the license_link "https://ai.google.dev/gemma/docs/gemma_4_license", with gated and private both false.

• https://huggingface.co/api/models/google/gemma-4-31B-it recorded 2026-09-29

  Hub record for google/gemma-4-31B-it carries the tag "license:apache-2.0" with the card license_link "https://ai.google.dev/gemma/docs/gemma_4_license"; the gated and private flags are both false, so the weights download without an access request.

• https://huggingface.co/api/models/google/gemma-4-E2B-it recorded 2026-09-29

  Hub record for google/gemma-4-E2B-it, tagged "base_model:finetune:google/gemma-4-E2B" and "license:apache-2.0", with the gated and private flags both false.

• https://huggingface.co/api/models/google/gemma-4-E4B-it recorded 2026-09-29

  Hub record for google/gemma-4-E4B-it, tagged "base_model:finetune:google/gemma-4-E4B" and "license:apache-2.0", with the gated and private flags both false.

• https://huggingface.co/google/gemma-4-31B-it/raw/main/README.md recorded 2026-09-29

  Front matter "license: apache-2.0" with a license_link to ai.google.dev; the card says "This release includes open-weights models in both pre-trained and instruction-tuned variants" and describes the pretraining mixture as "web documents, code, images, audio, with a cutoff date of January 2025" without releasing it. Searched for a training-pipeline or pretraining-code repository: none linked, only the google-gemma GitHub organization and inference snippets.

Adoption is measured as Hugging Face downloads summed across six declared checkpoints spanning three generations. Almost all of that traffic now goes to the two Gemma 4 checkpoints, with the Gemma 2 and 3 repositories seeing comparatively little use.

• https://huggingface.co/api/models/google/gemma-2-27b recorded 2026-08-13

  6,087 downloads in the trailing 30 days for google/gemma-2-27b

• https://huggingface.co/api/models/google/gemma-2-9b recorded 2026-08-13

  77,700 downloads in the trailing 30 days for google/gemma-2-9b

• https://huggingface.co/api/models/google/gemma-3-12b-pt recorded 2026-08-13

  34,590 downloads in the trailing 30 days for google/gemma-3-12b-pt

• https://huggingface.co/api/models/google/gemma-3-27b-pt recorded 2026-08-13

  10,455 downloads in the trailing 30 days for google/gemma-3-27b-pt

• https://huggingface.co/api/models/google/gemma-4-26B-A4B-it recorded 2026-08-13

  10,074,016 downloads in the trailing 30 days for google/gemma-4-26B-A4B-it

• https://huggingface.co/api/models/google/gemma-4-31B-it recorded 2026-08-13

  9,915,701 downloads in the trailing 30 days for google/gemma-4-31B-it

Gemma 4 31B scores 19 on Artificial Analysis's Intelligence Index, well behind the highest-scoring open-weight releases from DeepSeek and GLM. Its model card reports 85.2% on MMLU Pro, 84.3% on GPQA Diamond and 80.0% on LiveCodeBench v6, and on the Arena text leaderboard it trails more than a dozen openly licensed models, which places it near rather than at the frontier of open models.

• https://arena.ai/leaderboard/text recorded 2026-09-29

  Text leaderboard entry "gemma-4-31b" linking to "https://aistudio.google.com/app/prompts/new_chat?model=gemma-4-31b-it" from Google, license Apache 2.0, now rank 73 at a rating near 1453 among 409 listed models; fifteen openly licensed models rank above it, including glm-5.3-max, deepseek-v4.1-flash-max and kimi-k3-max.

• https://artificialanalysis.ai/models/gemma-4-31b recorded 2026-09-29

  Artificial Analysis page on Intelligence Index v4.3.2: "Gemma 4 31B (Reasoning) scores 19 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models (median: 8)".

• https://cloud.google.com/blog/products/ai-machine-learning/gemma-4-available-on-google-cloud recorded 2026-09-29

  Google Cloud availability post: "Built from the same research as Gemini 3 and released under a commercially permissive Apache 2.0 license". It carries no benchmark figures and no Arena ranking.

• https://huggingface.co/google/gemma-4-31B-it recorded 2026-09-29

  Rendered model card: "Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output"; the evaluation table, 31B column, lists MMLU Pro 85.2%, AIME 2026 (no tools) 89.2%, LiveCodeBench v6 80.0%, GPQA Diamond 84.3% and MMMU Pro 76.9%.