Video: why Gemma 4 was free
A recent video argued that giving away models like Google’s Gemma 4 is often a platform play—commoditising one layer to drive adoption of adjacent tooling, cloud primitives and developer workflows. The piece recommends asking which layer a proposal commoditises and which control surface it protects when evaluating architectural choices. (The real reason Google gave away Gemma 4
Google released Gemma 4 on March 31 under an Apache 2.0 license, making the model weights free to download, tune and deploy. (blog.google) Large language models are prediction engines: they guess the next token in a sequence, and “open weights” means developers can run that engine on their own hardware instead of only through a company’s paid application programming interface. Google says Gemma 4 comes in four sizes — E2B, E4B, 26B Mixture of Experts and 31B Dense — and is built for text, image and some audio inputs. (ai.google.dev) Google did not release Gemma 4 as a standalone product detached from its stack. In its launch post, the company said the models “complement our Gemini models,” and Google Cloud documents list Gemma in Vertex AI’s Model Garden, where customers can test, tune and deploy it with Google tooling. (blog.google) (cloud.google.com) That is the business logic behind the recent YouTube explainer: give away one layer, then make money on the surrounding layers. In software, companies often cut the price of the engine to sell the road, the fuel and the dashboard — in this case cloud compute, model hosting, fine-tuning pipelines and developer workflows. (youtube.com) (cloud.google.com) Google’s own materials point in that direction. The company promoted Gemma 4 for Android’s AI Core, Google AI Edge, Hugging Face, Kaggle and Vertex AI, tying the “free” model to distribution channels, hardware targets and managed services that Google already controls or partners on. (developers.googleblog.com) (huggingface.co) (blog.google) The license matters because earlier “open” artificial intelligence releases often came with narrower terms. Gemma 4’s model card says the release uses Apache 2.0, and Google’s developer docs say the weights permit responsible commercial use, which lowers legal friction for startups deciding whether to build on them. (ai.google.dev 1) (ai.google.dev 2) That does not mean Google stopped pushing proprietary models. Vertex AI’s model catalog places Gemma alongside Gemini, and Google’s launch post framed the pair as a combined offering rather than a replacement. (cloud.google.com) (blog.google) The video’s takeaway is less about one release than about how to read platform strategy. When a company makes a model free, the useful question is not only what got cheaper on March 31, but which layer of the stack the company still expects developers to use next. (youtube.com) (cloud.google.com)
