Model components / Base / pretrained models

Gemma 4

Google's latest open-model family (launched March 2026, built from the same research as Gemini 3): E2B/E4B edge variants, a 26B-A4B MoE, and a 31B dense model, with context windows up to 256K, native vision + audio, and 140+ languages. Notably released under Apache 2.0 - a shift from earlier Gemma generations' custom license.

Apache-2.0 (OSI) - unlike CodeGemma/Gemma 3, which use the Gemma Terms of Use. Training data and recipe not released. ~8.4M monthly downloads on gemma-4-31b-it. Benchmark claims (LMArena) are vendor-reported. Consolidated on 2026-07-29 from gemma-2, gemma-3; openness follows gemma-4, the release that currently governs. See sources/slug_aliases.yaml.

Openness

3 medium confidence

3.0

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. The tier reads 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 SKUs are Apache-2.0 and ungated, verified against the Hub on 2026-07-29. The current release governs, so a family that becomes more open is recorded as such rather than pinned to its worst historical release. Adoption 5 is carried from Gemma 2, which still out-downloads Gemma 4.

https://ai.google.dev/gemma/docs/releases recorded 2026-06-17
Gemma 4 sizes, release dates, Apache-2.0 license

Adoption

5 medium confidence

5.0

Among the most-downloaded open model families; the 27B-IT shows ~133.6k downloads/mo and the family runs widely via HF, Ollama, Kaggle, Google AI tooling (27 finetunes, 14 quantizations, 10 adapters listed on the base card). Family-level reach in the tens-of-millions; calibrated to the Gemma 3 anchor (A5).

https://huggingface.co/google/gemma-2-27b-it recorded 2026-06-04
~133.6k downloads/mo for Gemma 2 27B-IT
https://huggingface.co/google/gemma-2-27b recorded 2026-06-04
27B base 5,893 downloads/mo, 27 finetunes / 14 quantizations / 10 adapters, broad multi-platform distribution

Capability

5 medium confidence

5.0

Frontier-class for an open model; benchmark ranking is vendor-reported.

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

Unchanged since 2026-07-30 (last edited, not re-checked)