Gemma 4: What an Apache 2.0 Open Model Means for Your Business

In short: Gemma 4 is Google's family of open-weight AI models, released on April 2, 2026 under the Apache 2.0 license, which allows commercial use. For a business, the useful question is not whether it beats a cloud AI service on a leaderboard, but whether running a model you control fits your privacy, cost and language needs.

Key Takeaways

• Gemma 4 launched in four sizes (the model card now also lists a 12B), from small models for phones and laptops to a 31B model for a single powerful GPU.
• "Open-weight" means you can download the model and run it yourself; a closed API means the provider runs it and you send your data to them.
• Apache 2.0 is a permissive license that allows commercial use, but you should read the license text and the model card yourself.
• Benchmark rankings are Google's claims at launch. Test the model on your own documents before you commit.

What Is Gemma 4?

Gemma 4 is a family of open models from Google DeepMind. According to Google's announcement on April 2, 2026, it comes in four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture of Experts model and a 31B dense model. Google's model card also lists a 12B variant, so the family is five sizes in total.

• Context window: 128K on the two edge models and up to 256K on the larger ones.
• Inputs: all models handle images and video, and E2B and E4B add native audio input for speech recognition.
• Skills: Google highlights multi-step reasoning, function calling and structured JSON output for agent workflows, code generation and support for more than 140 languages.
• Where to get it: Google AI Studio, Hugging Face, Kaggle and Ollama, plus tools such as vLLM, llama.cpp and NVIDIA NIM.

Google also says the 31B model ranks third among open models on the Arena AI text leaderboard and the 26B model sixth, and that Gemma 4 "outcompetes models 20x its size." These are the vendor's own statements, so treat them as a starting point rather than a verdict.

Open-Weight Model vs. Closed API: What Is the Difference?

With a closed API, such as Google's own Gemini models, the provider runs the model on its servers. You send your prompts and data over the internet, pay per use and get updates automatically. Our guide to reading a frontier model release covers how to judge that kind of launch.

With an open-weight model, the trained model files are published and you can run them on your own hardware or your own cloud account. You control where the data goes, but you also take on the work of hosting, monitoring and updating. "Open-weight" is not the same as having the training data or the full training recipe, so it is a narrower kind of openness than the name might suggest.

What Does the Apache 2.0 License Allow?

Google describes Gemma 4 as released under Apache 2.0, a widely used permissive open-source license. In general terms, it allows you to use, modify and distribute the software, including in commercial products, as long as you keep the license and copyright notices. It also includes an express patent grant.

Two practical points. First, the license covers the model files as published, so check that the exact version you download carries the license you expect. Second, this is a general description, not legal advice: read the license text and the model card, and ask a lawyer if your product depends on the answer.

Should a Business Self-Host or Use an API?

There is no single right answer. These are the questions that usually decide it:

1. Privacy and data location. If prompts contain customer records, health data or contracts, running a model inside your own environment can simplify compliance conversations.
2. Cost shape. An API charges per use and scales down to zero. Self-hosting means paying for GPUs or rented capacity whether or not anyone is using them, which pays off mainly at steady, high volume.
3. Latency and offline use. The small models are designed to run on phones, laptops and edge devices, which matters where connectivity is unreliable.
4. Language coverage. Google lists support for more than 140 languages. If your users write in Nepali, Hindi or another South Asian language, test on real examples, because a language being supported does not mean it works well for your task.
5. Team capacity. Someone has to run, secure and update the model. If you have no one for that, an API is usually simpler.

A common pattern is to prototype with an API, then move a stable, high-volume or sensitive workload to a self-hosted open model once the requirements are clear.

What Should You Be Careful About?

• Benchmarks are vendor claims. Leaderboard positions change and may not reflect your task. Build a small test set from your own documents and compare models on it.
• You own the operations. Hosting, security patches, access control and monitoring are your responsibility.
• You own the safety layer. An open model gives you control, but also the job of filtering outputs, limiting what connected tools can do and logging what the system does. The same principles apply as for any AI tool, as we describe in our explainer on superintelligence and AI safety.
• Check the license and model card. Confirm the terms and any usage guidance before building on it.

The Short Version

Gemma 4 gives businesses a capable open-weight option under a permissive license, in sizes that range from phone to single-GPU server. Whether it is the right choice depends on your data, your volume and your team, not on a leaderboard. Start with a small test on your own material.

If you are weighing options, see our checklist for choosing an AI development company.

Related Insights

• Gemini 4 Argon: How to Read a Frontier Model Release
• What Is Superintelligence and Why Is It Called That?
• Why AI Agents Are Getting Their Own Infrastructure

Sources

• Google, Gemma 4: Byte for byte, the most capable open models, April 2, 2026
• The Apache Software Foundation, Apache License, Version 2.0