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