Gemma 4

Gemma 4 is a family of AI models from Google that anyone can download for free and run on their own computer, server or phone. Because the model runs on your own hardware, customer data does not have to go to an outside service. Google released it on 2 April 2026 under the Apache 2.0 licence, which also allows commercial use. All models read text, images and video.

What was in the family in April 2026?

Google DeepMind released Gemma 4 in four sizes. They are open-weight models: you can download the trained model, but Google does not publish the training data.

• E2B and E4B are meant for phones and small devices such as a Raspberry Pi. The E stands for effective: 2.3 and 4.5 billion parameters do the computing, and with their embedding tables, which are only used for lookups, the models have 5.1 and 8 billion.
• 31B is a dense model with 31 billion parameters.
• 26B A4B is a mixture-of-experts model: of its 26 billion parameters it activates 3.8 billion per step, according to Google.

All four are language models that answer in text and accept text, images and video. Only E2B and E4B also accept audio, for speech recognition and questions about speech. According to the Hugging Face launch post (archived on 26 April 2026), music and other sounds were not in the training data. The small models have a context window of 128,000 tokens, the large ones 256,000. Google says the models were trained on more than 140 languages and support function calling and structured JSON output, so they can be used in agents. In June 2026 Google added a fifth size, Gemma 4 12B, its first mid-sized model that also accepts audio.

How does it work?

Google says Gemma 4 is built from the same research and technology as Gemini 3. The Hugging Face post lists the main parts:

• The attention layers alternate between a local sliding window (512 tokens in the small models, 1,024 in the large ones) and the full context.
• Per-layer embeddings give each layer of the small models its own small embedding for every token.
• In the shared KV cache, the last layers reuse the keys and values of earlier layers, which saves memory with long inputs.
• E2B and E4B have an audio encoder, a conformer in the style of Google's USM, as in Gemma 3n.

According to Google, the unquantized 31B and 26B models fit on a single NVIDIA H100 with 80 GB of memory, and quantized versions run on consumer graphics cards. The weights are on Hugging Face, Kaggle and Ollama, and tools such as Transformers, llama.cpp, vLLM, MLX and LM Studio supported the models from day one. You can fine-tune them on your own data, for example with TRL or Unsloth.

How good was it in April 2026?

In the Hugging Face post, the instruction-tuned 31B model scores 85.2% on MMLU Pro and 80.0% on LiveCodeBench v6. Its estimated LMArena score for text is 1452; the mixture-of-experts model reaches an estimated 1441 with about 4 billion active parameters. At launch Google said the 31B model ranked third among open models on the Arena AI text leaderboard, and the 26B model sixth. Google and Hugging Face published these figures themselves.

What are the limits?

• Audio only works in the two small models, and only for speech.
• Tools lagged behind at launch. On 2 April Simon Willison found no way to use audio input in LM Studio or Ollama yet, and the 31B model repeated the same output in a loop in LM Studio.
• Google does not publish the training data. Even Hugging Face, which worked with Google on the launch, wrote that it did not know the training mix.

Example

Suppose an accounting firm wants to read scanned invoices without sending them to a cloud service. It runs the 26B model on a workstation with a graphics card, gives it each scan and asks for the invoice number, date and amount as JSON. Because the model runs on the firm's own machine, the invoices stay inside the office.