When Google released Gemini 3 Pro Late last year, it was a significant step forward for the company’s large, proprietary language models. Today, the company is bringing some of the same technology and research that made these models possible to the open source community with the release of its new family of open-weight Gemma 4 models.
Google offers four different versions of Gemma 4, differentiated by the number of settings offered. For cutting-edge devices, including smartphones, the company offers “efficient” 2 billion and 4 billion models. For more powerful machines, there are the 26 billion “Expert Blend” systems and the 31 billion “Dense” systems. For those unfamiliar, parameters are the parameters that a large language model can modify to generate output. Generally, models with more parameters will provide better answers than those with fewer, but running them also requires more powerful hardware.
With Gemma 4, Google claims to have succeeded in designing systems with “an unprecedented level of intelligence per parameter”. To support this claim, the company highlights the performance of the 31 billion and 26 billion variants of Gemma 4, which took third and sixth place in the market, respectively. Arena AI Text Leaderboardbeating models 20 times larger.
All models can process video and images, making them ideal for tasks like optical character recognition. The two smaller models are also capable of processing audio input and understanding speech. Furthermore, Google claims that the Gemma 4 family is capable of offline code generation, meaning you can use them to do vibrational coding without an internet connection. Google also trained the models in more than 140 languages.
Google launches the Gemma 4 family under a Apache 2.0 License. The company has made previous Gemma models available through its own Gemma License. The move will give people greater freedom to modify the new systems according to their needs.
“This open source license provides the basis for complete developer flexibility and digital sovereignty; giving you full control over your data, infrastructure and models. Google said. “It lets you build freely and deploy securely in any environment, whether on-premises or in the cloud. »
If you’d like to try one of the systems for yourself, model weights are available through Hugging Face, Kaggle, and Ollama.
This article was originally published on Engadget at https://www.engadget.com/ai/google-releases-gemma-4-a-family-of-open-models-built-off-of-gemini-3-160000332.html?src=rss
