Google has released a new family of large language models, Gemma 4, based on Gemini 3 technologies. Gemma 4 is available under the Apache license in versions with 2.3, 4.5, 25.2, and 30.7 billion parameters (E2B, E4B, 31B, and 26B A4B). The E2B and E4B versions are suitable for mobile devices, Internet of Things (IoT) systems, and Raspberry Pi boards, while the other versions are suitable for workstations and systems with consumer GPUs. The context size considered by the model is 128,000 tokens for the E2B and E4B models, and 256,000 tokens for the 31B and 26B A4B models.
The models are multilingual and multimodal: out of the box, they support 35 languages (over 140 languages were used during training), and they can process text and images (the E2B and E4B models also support audio processing). The 26B A4B model is based on a Mixture-of-Experts (MoE) architecture, where the model is divided into a series of expert networks (only 3.8 billion parameters may be used for response generation, but the speed is significantly higher than traditional large models), while the other versions use a classic monolithic architecture.
The models support reasoning and customizable thinking modes, and they support System Role for processing instructions (rules, constraints) separately from the data. The models can be used for coding, object recognition in images, frame-by-frame video analysis, parsing documents and PDFs, optical character recognition (OCR) for printed and handwritten text, speech recognition, and language translation. They can also be used as autonomous agents, interacting with various tools and APIs.
In most tests, the Gemma 4 series models significantly outperformed the Gemma 3 model with 27 billion parameters. Gemma 4 is supported for use with tools and libraries such as LiteRT-LM, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM and NeMo, LM Studio, Unsloth, SGLang, Cactus, Basetan, MaxText, Tunix, and Keras.
Source: opennet.ru
