Google introduces Gemma 4, a new family of artificial intelligence models that emerges from the same research line as Gemini 3, but with a key difference: Gemma 4 is completely open source and suitable for commercial use. Unlike previous versions, which imposed restrictions and were not fully open, Gemma 4 adopts the Apache 2.0 license, allowing unrestricted use, in a clear move to compete with Meta's Llama models, which also use this license. The Gemma 4 family is launched in four sizes, designed to cover everything from mobile devices to advanced workstations. All models are multimodal, capable of processing video and images natively and trained in over 140 languages. According to Google, these models have been optimized to work efficiently across a wide range of hardware, from billions of Android devices to GPUs in laptops and workstations for developers. One of the main innovations of Gemma 4 is its focus on agent-based workflows. The models offer native support for function calls, structured JSON output, and system instructions, making it easier to create autonomous agents capable of executing complex logic and interacting with external APIs locally. In terms of performance, Google claims that the 31B Gemma 4 model ranks third among open models in the Arena AI ranking, while the 26B model ranks sixth, outperforming much larger models. The unquantized weights of the 26B and 31B models can be hosted on a single 80 GB NVIDIA H100 GPU. For local development, the 26B Mixture of Experts (MoE) model is optimized to reduce latency, activating only 3.8 billion parameters during inference, allowing for extremely fast token generation, ideal for coding assistants on consumer graphics cards. Multimodality is another strong point of Gemma 4. Following the line of Gemma 3n, the entire family processes high-resolution video and images, and the Edge E2B and E4B models add native audio input for voice recognition with minimal latency. These models offer context windows of 128 KB on edge devices and up to 256 KB on larger models. Gemma 4 is compatible with platforms like Hugging Face, Ollama, and vLLM, and is optimized for hardware from NVIDIA, AMD, Qualcomm, and MediaTek. Additionally, the models are ready for prototyping in the AICore preview version, ensuring compatibility with the future Gemini Nano 4. Source: muycomputerpro.com