Gemma 4

weights

open(Apache-2.0)

data

closed

code

closed

license

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

Apache-2.0 weights, no open training data and no training code, so open weights rather than a fully open pipeline. It scores 3 because Gemma 4 relicensed: Gemma 2 and 3 shipped under the non-OSI Gemma License with a Prohibited Use Policy and scored 2, while all six Gemma 4 variants are Apache-2.0 and ungated on Hugging Face. The current release governs, so a family that becomes more open is recorded as such rather than pinned to its most restrictive historical release. The Apache-2.0 claim rests on the Gemma 4 model card front matter and on ai.google.dev/gemma/docs/gemma_4_license, which redirects to ai.google.dev/gemma/apache_2; the releases page carries no Apache statement for the models, only a site footer covering code samples. The card documents the pretraining mixture without releasing it and links no training pipeline, so data and code stay closed.

• https://huggingface.co/google/gemma-4-31B-it/raw/main/README.md recorded 2026-08-13

  front matter `license: apache-2.0` with `license_link` to ai.google.dev/gemma/docs/gemma_4_license; a Model Data section describing the pretraining mixture as "web documents, code, images, audio, with a cutoff date of January 2025" without releasing it; no training-pipeline repository linked.

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

  card license `apache-2.0`, `gated: false`, `private: false`, 9,915,701 downloads. The Gemma 4 SKUs are ungated, unlike the Gemma 2 and 3 repos which are still `gated: manual`.

• https://ai.google.dev/gemma/docs/gemma_4_license recorded 2026-08-13

  The Gemma 4 license URL redirects to ai.google.dev/gemma/apache_2, which is what settles the relicensing rather than a Gemma Terms of Use page.

• https://ai.google.dev/gemma/docs/releases recorded 2026-08-13

  Gemma 4 sizes and release dates. It carries no Apache-2.0 statement for the models - the only Apache text on the page is the site footer covering code samples.

20,118,549 downloads in the trailing 30 days across the six declared artifacts (google/gemma-2-27b 6,087; google/gemma-2-9b 77,700; google/gemma-3-27b-pt 10,455; google/gemma-3-12b-pt 34,590; google/gemma-4-31B-it 9,915,701; google/gemma-4-26B-A4B-it 10,074,016), which bands at level 5 (>10M) on the model adoption scale. The two Gemma 4 checkpoints are 99.4% of that total and the Gemma 2 repositories are now a rounding error, so the family's reach rests on the current generation rather than an older one.

• 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-3-27b-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-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-4-31B-it recorded 2026-08-13

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

• 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

Frontier-class for an open model. The 5 rests on the evaluation table Google itself publishes on the gemma-4-31B-it model card - MMLU Pro 85.2%, GPQA Diamond 84.3%, AIME 89.2% without tools, LiveCodeBench v6 80.0%, MMMU Pro 76.9%. Audio is native only on E2B, E4B and 12B, so the flagship 31B is multimodal on text and image but not audio. No top-tier arena placement backs the score: lmarena.ai now redirects to arena.ai, where gemma-4-31b sits at rank 60 of 391 with a score of 1451 +/-8, roughly sixth among openly-licensed entries and behind glm-5.2-max, mimo-v2.5-pro, glm-5.1, glm-5 and hy3 - a strong open placement rather than a leading one.

• https://huggingface.co/google/gemma-4-31B-it recorded 2026-08-14

  "Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output"; evaluation table, 31B column - MMLU Pro 85.2%, AIME 2026 (no tools) 89.2%, LiveCodeBench v6 80.0%, GPQA Diamond 84.3%, MMMLU 88.4%, MMMU Pro 76.9%, MATH-Vision 85.6%. Neither "Arena" nor "LMArena" appears on the card.

• https://arena.ai/leaderboard/text recorded 2026-08-14

  successor to lmarena.ai, which 301s here - "gemma-4-31b", Google, Apache 2.0, rank 60 at 1451 +/-8, with gemma-4-26b-a4b at rank 80; several openly-licensed models rank above it. Does not support a top-tier Arena claim.

• https://cloud.google.com/blog/products/ai-machine-learning/gemma-4-available-on-google-cloud recorded 2026-06-17

  Gemma 4 availability on Google Cloud. Carries no multimodal statement and no Arena ranking, so it does not establish either half of what this axis cited it for.