انتقل إلى المحتوى
العودة إلى المدونة
4 دقيقة قراءةبالإنجليزية

Google Gemma 4 Brings Capable AI to Your Own Hardware

Google releases Gemma 4 under the Apache 2.0 licence, with open models built for reasoning and agentic work that outperform models 20 times their size.

  • open source AI
  • Google Gemma
  • on-device AI
  • agentic AI

Google has released Gemma 4, a family of open AI models built from the same research as Gemini 3, under the Apache 2.0 licence. For organisations that would rather keep data on their own servers than send it to a cloud API, this is one of the most useful open releases so far.

Google Gemma 4 open AI model announcement
Google Gemma 4 open AI model announcement

What Google released

The headline result: the 31B dense model ranks third on Arena AI's text leaderboard, and the 26B mixture of experts model ranks sixth. Both outperform models 20 times their size.

Google has designed Gemma 4 for agentic workflows and multi-step reasoning. Earlier open models handled simple questions well enough but struggled when they had to break a task into steps, carry them out, and adjust based on what happened along the way. That kind of sequential planning is what most real AI applications need, and it is the focus of this release.

There are four models:

  • Gemma 4 31B Dense: the most capable, with a 256K context window, for cloud servers and high-end workstations
  • Gemma 4 26B MoE: a mixture of experts design with similar quality and more efficient inference
  • Gemma 4 E4B (Effective 4B): built for edge devices, with a 128K context window and native audio input
  • Gemma 4 E2B (Effective 2B): small enough to run on smartphones, Raspberry Pi and Jetson Nano

All four read images and video natively, including text in images (OCR) and charts. The two edge models also handle speech recognition and understanding. The models were trained natively on more than 140 languages.

Why the licence matters

Apache 2.0 removes the restrictions that made earlier Gemma versions awkward to use commercially. You control the data, the infrastructure and where the model runs. For organisations in the Gulf with data sovereignty requirements, that control is often a precondition for using a model at all.

The licence also puts pressure on other open model providers. Meta's Llama models still carry a more restrictive licence. Alibaba's Qwen family and Zhipu's GLM models are strong alternatives, but Gemma 4 combines high benchmark results with Google's enterprise support. For teams that need reasoning and agentic capability with full commercial rights, it is now a natural first model to evaluate.

Running it on devices

Google worked with Pixel, Qualcomm and MediaTek on the edge models and claims near-zero latency on supported hardware. Running a model on the device addresses two problems at once. Responses do not wait for a round trip to a cloud server, and sensitive information never leaves the device or the internal network. The 128K context window on the edge models means fairly long documents can be processed without being cut short.

The weights are available on Hugging Face, Kaggle and Ollama, so they fit into existing machine learning pipelines. Google AI Edge Gallery provides builds tuned for mobile and embedded devices.

Where it could fit in UAE organisations

A few uses stand out, each depending on the model staying inside your own environment:

Arabic documents. Native training on more than 140 languages, combined with OCR, makes Gemma 4 a candidate for processing Arabic text and scanned documents. Test it on your own documents before relying on it.

Financial compliance. Banks and financial institutions could run the model on their own premises to analyse documents and support regulatory work without sending data to an outside AI provider. The 256K context window can take long regulatory texts in one pass.

Healthcare. Local inference allows patient-facing assistants that keep medical information inside the hospital network, with step-by-step reasoning that staff can check.

Government services. Agentic workflows could help with multi-step citizen requests such as document checks and eligibility screening. In a government setting these should keep a human reviewing the outcome.

The gap between open and closed models keeps narrowing. When an open model beats much larger alternatives, paying for an API-only service becomes harder to justify for many tasks.

If your team wants to try it, start with the E4B model on a real, low-risk internal task with non-sensitive data. It is quick to set up, and because the model family shares one design, the prompts and workflows you build should carry over to the 26B or 31B models when you need more capability.