Google has released Gemma 4, its most capable open-weight AI model family, under a fully permissive Apache 2.0 license, marking a significant departure from the custom licensing restrictions that previously limited enterprise adoption of its earlier Gemma releases. Released on April 2, 2026, Gemma 4 is built from the same research foundation as Gemini 3 and arrives as a structured family of four models designed to run everywhere from smartphones to data centers .

For the past two years, enterprises evaluating Google's open-weight models faced a familiar frustration: strong performance paired with licensing ambiguity. Custom clauses, usage restrictions, and mutable terms kept legal teams cautious and often steered organizations toward competitors like Mistral AI or Alibaba's Qwen. With Gemma 4, that calculus changes entirely .

What Makes Gemma 4 Different From Previous Google Models?

The shift from Google's proprietary licensing to Apache 2.0 represents more than a legal technicality. Apache 2.0 is the same permissive framework used across much of the open-source ecosystem, meaning no custom clauses, no "harmful use" carve-outs, and no ambiguity around redistribution or commercial deployment . This timing is particularly notable because it arrives just as Chinese AI providers like Tencent and ByteDance are moving in the opposite direction, abandoning open-weight models in favor of proprietary systems .

Gemma 4 comes in four distinct configurations, each optimized for different deployment scenarios:

• Dense Models: A 31-parameter dense model and a 26-parameter Mixture-of-Experts (MoE) model, both supporting text and image input with 256,000-token context windows, roughly equivalent to processing 200,000 words at once
• Edge Models: The E2B and E4B variants designed for phones, embedded systems, and laptops, supporting text, image, and audio with 128,000-token context windows
• Multimodal Capabilities: Native support for text, image, and video inputs across all sizes, with audio processing built directly into edge models for on-device speech recognition and translation
• Function Calling: Built-in agentic function calling at the architectural level, not layered through prompting, improving reliability for autonomous agent systems

The MoE model represents the most architecturally interesting choice. Instead of using a few large experts, Google deployed 128 small experts, activating eight per token plus a shared expert that's always active. The result is a model that performs competitively with dense 27 to 31-parameter models while operating at roughly 4-parameter inference cost . In practical terms, this means fewer graphics processing units (GPUs), lower latency, and reduced serving costs, turning what could be an interesting research model into a genuine production candidate.

How Does Gemma 4 Perform on Real-World Tasks?

Gemma 4 demonstrates strong reasoning and coding capabilities across multiple benchmarks. The dense 31-parameter model scored 89.2% on the AIME 2026 mathematics benchmark and 80% on LiveCodeBench v6, a coding evaluation suite . The model achieved a Codeforces rating of 2,150, a competitive programming benchmark that reflects practical problem-solving ability. Even the smaller edge models outperform expectations, delivering results that surpass earlier large models despite dramatically smaller footprints .

Beyond raw benchmark numbers, Gemma 4 was pretrained on 140 languages with support for 35 languages out of the box, making it genuinely useful for multinational enterprises . The long context windows, up to 256,000 tokens on the larger models, enable developers to reason across entire codebases, lengthy legal documents, or multi-session conversation histories without losing context .

Steps to Evaluate Gemma 4 for Your Organization

• Audit Current Deployment Constraints: Review whether your existing AI applications are limited by cloud costs, data privacy requirements, or latency sensitivity. Gemma 4's on-device capabilities and permissive licensing may unlock projects previously rejected due to these constraints
• Test on a Specific Use Case: Identify one production use case in your operations, such as in-store analytics, branch customer service, or in-vehicle systems, and conduct a 30-day feasibility study using Gemma 4 to compare performance and cost against your current solution
• Evaluate Licensing and Compliance: Work with your legal team to confirm that Apache 2.0 licensing aligns with your commercial deployment plans, redistribution needs, and regulatory requirements, particularly if you operate in regulated industries
• Compare Total Cost of Ownership: Calculate infrastructure costs for Gemma 4 deployment, including GPU requirements for the MoE model versus dense alternatives, and compare against your current cloud-based AI spending

Where Can You Deploy Gemma 4 Right Now?

Microsoft Foundry, the company's unified AI platform, now offers Gemma 4 alongside models from OpenAI, Anthropic, and over 11,000 other models under a single control plane . This means Azure customers can discover, evaluate, and deploy Gemma 4 inside their Azure environment with the same network policies, identity controls, and audit processes they rely on for every other workload .

Deployment options include managed online endpoints that handle serving, scaling, and monitoring without manual infrastructure setup, serverless deployment with Azure Container Apps for cost-sensitive applications, and Foundry Local, which lets developers run optimized Hugging Face models directly on their own hardware using the same model catalog and software development kit (SDK) patterns as cloud deployments .

Google also introduced serverless GPU support via Cloud Run, introducing a deployment model where inference scales to zero when idle, potentially reshaping cost structures for internal tools and lower-traffic applications .

What Does This Mean for Enterprise AI Strategy?

The release of Gemma 4 under Apache 2.0 signals a fundamental shift in how enterprises can approach open-weight AI. For organizations that had been waiting for Google to align with industry norms around open licensing, the wait is over. The combination of strong performance, native multimodality, long context windows, built-in function calling, and fully permissive licensing in a single family that scales from phones to cloud deployments represents a genuine turning point .

For the first time, evaluating a Google open model doesn't start with a licensing discussion; it starts with capability . This removes a significant friction point that previously steered enterprises toward alternatives. The timing also matters: as Chinese AI providers pivot away from open-weight models toward proprietary systems, Google is moving in the opposite direction, opening up its most capable release yet .

The practical implications are substantial. Organizations can now reduce cloud costs through on-device deployment, address data sovereignty concerns without licensing ambiguity, and build agentic systems with reliable function calling rather than prompt-engineered workarounds. For retail, banking, automotive, and other industries with privacy-sensitive or latency-critical requirements, Gemma 4 offers a tangible path forward that wasn't available with previous Google open models .