Releasing model weights to the public isn't a deployment - it's closer to a product launch with no way to push updates afterward. Google DeepMind shared a look at what went into shipping Gemma 4, their latest open model series, and the coordination required goes well beyond the research team.
When you ship an API model, you maintain full control of inference. Behavior you don't like can be adjusted server-side. Abuse can be rate-limited. Safety filters can be updated without touching the underlying model. Open weights flip all of that. Once the file is public, it's permanent - developers will download it, fine-tune it (meaning: train it further on their own data to specialize its behavior), and embed it in production systems. Whatever's baked into those weights at release time is what ships.
That constraint shapes the entire launch process. Safety evaluations have to be thorough rather than iterative, because post-launch adjustments require releasing an entirely new model. Developer tooling needs to work before launch, not after. Licensing terms need to reflect what uses Google is and isn't comfortable enabling - and those terms stick.
The Gemma series has been Google's main open model line since 2024, positioned for developers who want full control over their inference stack - no API keys, no per-token costs, no external dependencies. Gemma 4 continues that trajectory, arriving as open models have become increasingly competitive with cloud API alternatives for practical tasks. Google's willingness to document the launch process reflects how seriously they're treating the open model space as a distinct product category, not just a research side project.