Google has taken a significant step forward in its open AI strategy with the launch of Gemma 4, a new family of open-weight models that aims to combine high performance, hardware efficiency, and a truly open license for commercial use. Built on the same technological foundation as Gemini 3, this line targets both large enterprises and developers who want to deploy advanced AI without relying entirely on closed cloud services.
Far from being just another experimental model, Gemma 4 arrives as a complete offering of four variants capable of running on mobile devices, edge devices, personal computers, and servers with high-performance GPUs. Google's strategy focuses on providing more intelligence per parameter, reducing infrastructure costs, and simultaneously giving the community and businesses the flexibility to adapt the models to their own needs.
A family of four models designed to cover everything from mobile to data center
The Gemma 4 family is organized into four main sizes: E2B, E4B, 26B MoE, and 31B Dense . The first two are geared towards edge execution, while the 26.000 billion and 31.000 billion parameter models target high-performance workstations, including high-performance laptops and server environments.
The Effective 2B (E2B) and Effective 4B (E4B) variants have been specifically designed for resource-constrained devices, such as Android phones, IoT boards, and embedded systems like Raspberry Pi or hardware from manufacturers such as Qualcomm and MediaTek. Their goal is to maintain good processing power and multimodality while minimizing memory and battery consumption and latency.
Above all, the 26B model with Mixture of Experts (MoE) architecture is optimized to minimize response time: during inference it only activates around 3,8 billion parameters, allowing high-speed token generation on developer hardware or consumer GPUs, including custom AI chips , ideal for local programming assistants and development tools.
At the top end is Gemma 4 31B Dense , the task-oriented dense variant that prioritizes quality and consistency over speed. This model has already secured top positions in open model rankings such as the Arena AI text leaderboard, competing with systems twenty times its size in terms of parameters.
This combination of lightweight and heavyweight models allows Gemma 4 to cover everything from everyday mobile uses to critical enterprise workflows , giving systems architects the flexibility to choose between speed of inference or depth of analysis depending on the project.
Extended multimodality and long context windows
One of the strengths of the new family is its ability to work natively with multiple content types . All Gemma 4 models can process text and images, supporting different resolutions and aspect ratios, which facilitates use cases such as analyzing scanned documents, visually understanding interfaces, or generating descriptions.
Furthermore, the E2B and E4B versions extend multimodality to video and audio , enabling them to handle low-latency speech recognition, video clip analysis, or augmented reality applications directly on the device. In mobile or IoT scenarios, this ability to run vision and audio without constantly relying on the cloud reduces connectivity issues and improves privacy.
Regarding the handling of large amounts of data, the Gemma 4 family introduces context windows of up to 256.000 tokens in the larger models . The edge device variants offer contexts of 128K, while the 26B and 31B versions reach 256K tokens. This allows, for example, loading entire code repositories, large document databases, or very long conversation histories in a single query.
This broad context is particularly useful for offline code generation, automated technical support, or legal document analysis —areas of special relevance in European companies subject to strict regulations and which often need to keep information within their own systems.
Along with multimodality and expanded context, Google highlights Gemma 4's support for over 140 languages . This broad linguistic coverage makes it an attractive option for companies with a global presence, European public administrations, or startups that want to launch multilingual products without relying on multiple different models.
Autonomous agents, JSON, and function calls: Gemma 4 agentic flow-oriented
Gemma 4 goes beyond traditional text generation. The entire family has been designed with a clear focus on agent-based workflows , an increasingly relevant trend in enterprise and software development environments.
The models include native support for function calling , allowing the system to invoke external APIs or specific tools in a controlled manner. They also offer structured JSON output, facilitating integration with applications that require formatted responses for consumption by other services or microservices.
Another key aspect is compatibility with native system instructions , which allow for precise definition of the "system" role and the establishment of clear rules for the model's behavior. This capability is especially useful when building autonomous agents that manage customer service, automate internal processes, or coordinate various tools within a company.
According to Google Cloud representatives, enterprise AI demands models capable of executing complex logic while keeping data within secure environments . In this regard, Gemma 4's agentic approach combines with on-premises and controlled cloud deployment options to reduce risks and increase control over where and how data is processed.
The company accompanies these models with an Agent Development Kit (ADK) , a modular framework designed to accelerate agent design, and with support for running intensive workloads serverless in Cloud Run on NVIDIA RTX PRO 6000 GPUs (Blackwell), which reduces the initial investment needed to experiment with complex agents.
Apache License 2.0 and digital sovereignty: implications for Europe and Spain
One of the most significant changes compared to previous generations of Gemma is the license. For the first time, Gemma 4 is distributed under Apache 2.0, a fully permissive open license that allows commercial use without additional specific restrictions from Google.
In previous versions, the terms of use included conditions that raised concerns among corporate legal teams, especially in large companies and public administrations. With Apache 2.0, Google places Gemma 4 in the same licensing category as other leading open models like Llama , facilitating its adoption in production projects without the need for individual negotiations.
This decision has a clear European interpretation. The combination of an open model, compatibility with over 140 languages, and sovereign deployment options aligns with data residency regulations and the ongoing discussions surrounding the European Union's AI Regulation. Spanish and European companies can integrate Gemma 4 into their solutions, maintaining greater control over where data is stored and processed.
Google anticipates making Gemma 4 available in sovereign cloud environments and air-gapped configurations , as well as on-premises installations. For regulated sectors such as banking, healthcare, energy, and public administration, this opens the door to leveraging advanced AI without needing to send sensitive information to shared infrastructures outside the European area.
The license's flexibility also encourages the creation of local and specialized variants . Examples have already been seen in the past, such as models adapted to specific languages ​​and contexts (for example, BgGPT in Bulgaria or medical applications in North American universities), and Google expects Gemma 4 to strengthen this ecosystem, which some refer to as a "Gemmaverse" with tens of thousands of community variants.
Google Cloud integration, local execution, and required hardware
Beyond opening up the model, Google has prepared a supporting infrastructure centered on Vertex AI and Google Kubernetes Engine (GKE) . Through these services, organizations can provision tailored resources, scale inference workloads, and adjust deployment to their security and compliance requirements.
In Vertex AI, Gemma 4 is integrated as part of the model catalog, allowing technical teams to test, fine-tune, and deploy custom variants while maintaining control over computing resources. The combination with GKE enables dynamic scaling, adapting the number of inference service replicas to actual demand.
An important detail for mid-sized businesses is that the bfloat16 weights of the 26B and 31B models fit on a single 80GB NVIDIA H100 GPU . This significantly reduces the minimum investment required to access high-end models, compared to alternatives that require multiple GPUs in parallel.
At the same time, Gemma 4 is optimized to run on diverse hardware, from consumer GPUs to mobile solutions with 5G M2M connectivity . The E2B and E4B models leverage techniques such as Per-Layer Embeddings (PLE) to maximize per-parameter efficiency, allowing them to run on phones, Raspberry Pi, or edge devices with very low latency.
Compatibility also extends to ecosystems such as Hugging Face, Ollama, vLLM, LM Studio, and llama.cpp , as well as Google development platforms like AI Studio and AICore (for Android prototyping). This makes it easier for both independent developers and corporate teams to integrate Gemma 4 into their existing workflows without having to start from scratch.
Potential uses in business, education, and the public sector
Gemma 4's capabilities allow for the deployment of a wide range of practical applications that go beyond traditional chatbots. In the enterprise environment, the models can serve as the foundation for internal virtual assistants that answer questions about corporate documentation, generate executive summaries, or automate repetitive tasks in multiple languages.
In the field of programming, the combination of wide context windows, code generation, and low latency makes Gemma 4 suitable for local development assistants, automatic code review, or tools that analyze entire repositories in a single pass, keeping the code within the company's own infrastructure.
In education, Gemma 4 could be used to create personalized tutors that adapt content to the student's level, generate summaries of complex texts, or explain images and graphics, something especially useful for students with specific accessibility needs.
For the public sector and administrations in Spain and Europe, the possibility of deploying these models in controlled environments, with data residing in European territory , opens up options in citizen services, file analysis or automation of procedures, provided that they are integrated with the guarantees of transparency and human supervision required by the regulation.
In sectors such as manufacturing, precision agriculture, and infrastructure management, running data locally on edge computing devices allows for real-time data analysis without relying on a constant cloud connection . This reduces transmission costs, improves response times, and minimizes the exposure of sensitive data to external networks.
Local AI, costs, and the gap between open and proprietary models
The launch of Gemma 4 reflects a clear trend in the industry: the priority is no longer just who has the largest model, but who achieves the best balance between capacity, cost, and ease of deployment . Google emphasizes the idea of ​​"parameter intelligence" as the central metric for this new generation.
The ability to run advanced models locally, without always relying on large cloud services, points to a shift in how products and services are designed . For many everyday tasks—summarizing text, creating a reminder, processing a simple image—it doesn't make much sense to send data to massive remote models if it can be handled on the device itself.
Even so, Gemma 4 is not intended to replace Google's proprietary models, but rather to complement them . The company maintains Gemini as its most advanced and closed layer, reserved for use cases where maximum capacity is paramount. Gemma 4 sits a step below in terms of technological edge, but gains ground in openness, flexibility, and cost control.
For IT departments, this presents an increasingly visible choice: closed models, with greater ease of use but less control, versus open models that require more active infrastructure management in exchange for total sovereignty and greater economic optimization in the medium term.
In this context, the competitiveness of Spanish and European companies in the field of AI may depend, to a large extent, on their ability to integrate open models such as Gemma 4 into their critical processes , combining them when necessary with proprietary services and always ensuring compliance with data protection regulations and future European regulations on artificial intelligence.
With Gemma 4, Google consolidates a firm commitment to efficient open models, capable of running on accessible hardware, adapting to different regulatory frameworks, and serving as a basis for a new generation of local agents and applications; those who know how to take advantage of this combination of openness, performance, and control will have an advantage when building sustainable AI solutions aligned with the demands of Europe.
