Gemma 4 26B A4B

Gemma 4 26B A4B is an open-weight language model from Google DeepMind, released on April 2, 2026.

Developer: Google DeepMind
Released: April 2, 2026
Total parameters: 25.2B
Active parameters per token: 3.8B
Architecture: MoE
Context window: 256K
Input: Text, image
License: Apache 2.0
Commercial use: Yes
Official weights, GB: 51.6 GB
Hugging Face repository: huggingface.co
Checked on: September 29, 2026

Gemma 4 26B A4B is a mixture-of-experts model with 25.2B total parameters, of which 3.8B are active for each token. Google DeepMind lists a context window of 256K tokens. The model accepts text and image input. The weights are released under Apache 2.0, which allows commercial use. The official weights on Hugging Face take about 51.6 GB in BF16. Running the model needs at least that much memory across GPUs and system RAM, plus room for the context cache; quantized versions need less.

Total parameters, billions: 25.2B
Active parameters, billions: 3.8B
Context window, tokens: 256K
License type: Apache 2.0
License conditions: Apache 2.0: keep the license and notices; no limits on field of use or number of users. Gemma 4 is the first Gemma generation under Apache 2.0; earlier versions used the Gemma Terms of Use.
Weights precision: BF16
Official quantized or alternative versions: google/gemma-4-26B-A4B-it-qat-q4_0-gguf; google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
License text: ai.google.dev
Sources: Page URL
Hugging Face model card https://huggingface.co/google/gemma-4-31B-it/raw/main/README.md
blog.google announcement https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
opensource.googleblog.com announcement https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html
Hugging Face API https://huggingface.co/api/models?author=google&search=gemma-4
Hugging Face weights index https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/model.safetensors.index.json

Page	URL
Hugging Face model card	https://huggingface.co/google/gemma-4-31B-it/raw/main/README.md
blog.google announcement	https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
opensource.googleblog.com announcement	https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html
Hugging Face API	https://huggingface.co/api/models?author=google&search=gemma-4
Hugging Face weights index	https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/model.safetensors.index.json

Questions

What license does Gemma 4 26B A4B use?

The weights are released under Apache 2.0, which allows commercial use. Gemma 4 is the first Gemma generation under Apache 2.0; earlier versions used the Gemma Terms of Use.

Can I use Gemma 4 26B A4B commercially?

Yes. Keep the license and notices; there are no limits on field of use or number of users.

How much memory does Gemma 4 26B A4B need?

The official Gemma 4 26B A4B weights take about 51.6 GB (BF16). All 25.2B parameters must be loaded even though only 3.8B are active per token; Google’s quantization-aware 4-bit versions need less. The memory calculator checks which setups fit.

What does A4B mean in Gemma 4 26B A4B?

It is a mixture-of-experts model: of its 25.2B total parameters, 3.8B are active for each token, which decides speed.

Sources

Every figure comes from the developer’s model card, license text and official announcements, and from the model’s Hugging Face repository metadata. Each model page lists its sources and the date they were checked; the register is rechecked every week.