# Gemma 4 — weights, license and what is actually open

Google’s Gemma 4 generation spans edge and workstation sizes; this atlas record uses the 26B mixture-of-experts checkpoint as its reference point.

Canonical: https://brightaifuture.com/open-models/gemma
Format: model-family
Source publication: 2026-04-02
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. Arena placement cited by Google is a third-party leaderboard result, but this record does not independently reproduce it.

## Documented release

Gemma 4 26B-A4B

Organization: Google DeepMind

Initial Gemma 4 model release: 2026-03-31 (day precision). Source: https://ai.google.dev/gemma/docs/releases

Google Gemma 4 announcement: 2026-04-02 (day precision). Source: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/

The 26B-A4B reference checkpoint was in the initial March 31 release; later Gemma 4 updates are separate events.

Parameters: 26B total / 4B active (reference checkpoint)

Architecture: Hybrid local/global-attention multimodal transformer; the 26B reference variant uses mixture-of-experts routing.

Modalities: text, image

Context: Up to 256K tokens on the 26B reference variant

## License and commercial use

Apache-2.0

Permitted by Apache-2.0.

## What is open

Weights: available. Pre-trained and instruction-tuned weights are released. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Architecture: available. The official card describes the dense and MoE variants and attention design. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Inference code: available. Google documents supported runtimes and deployment routes. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Training code: not-established. The complete code used for the original training runs was not confirmed. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Training recipe: partial. The model card documents methods but does not provide a full reproduction recipe. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Data information: partial. Data categories and filtering are described at a high level. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Training data: not-established. No sufficiently specific public artifact was confirmed in this review. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Evaluation: partial. Official benchmark methodology and results are published; full independent reproduction is separate. Source: https://ai.google.dev/gemma/docs/core/model_card_4

Commercial use: available. Gemma 4 is released under Apache-2.0. Source: https://ai.google.dev/gemma/docs/core/model_card_4

## Uses and strengths described in sources

Reasoning and tool use

Vision input

Wide size range

Official QAT checkpoints

## Hardware and quantization

Google documents a 12B variant running in 16 GB; requirements for the 26B-A4B variant depend on format and context.

Official quantization-aware-trained checkpoints

## Independent evidence

Arena placement cited by Google is a third-party leaderboard result, but this record does not independently reproduce it.

## Limitations

Audio is limited to the E2B, E4B, and 12B variants

Training data is described but not released

## Provenance and history

{}

## Original sources

- [Gemma release log](https://ai.google.dev/gemma/docs/releases)
- [Gemma 4: Byte for byte, the most capable open models](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/)
- [Gemma 4 model card](https://ai.google.dev/gemma/docs/core/model_card_4)

## Continue exploring

- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence thread](https://brightaifuture.com/threads/open)
- [Combining image, audio, video, and text at the edge](https://brightaifuture.com/discoveries/gemma-3n-edge-assistance)