Pretty much what everyone else said. It's effectively a trend towards models becoming less of a proprietary product and more of a rudimentary scientific discovery. The LLM itself isn't really the product by itself anymore, now companies are offering services around their flagship LLMs. Google Gemini isn't "just" an LLM, it's the system around their internal flagship LLM. Any research done with LLMs that don't directly contribute to their proprietary services are sheer byproduct, just another part of all the other code they publish with papers when they discover a new algorithm. As for why they publicize it, as others said, good will, R&D, free publicity, stuff like that. Pretty good stuff to get out of a sheer byproduct.

This could potentially be why OpenAI seems to be falling behind Google - now that creating a state-of-the-art AI agent has shifted from scientific discovery with LLMs to more service building, heavily shifting from what OpenAI excelled at into what Google excels at. As someone else mentioned here about moats, companies could be trying to drown out the competition (especially OpenAI, the king of the hill) with free alternatives that might not be as good as their flagship services (like Gemini being backed by integrating with all of Google's services, both public and internal) but are just as serviceable for rudimentary personal assistants and automation tasks, ones that, even if they don't go to their own business, at least it doesn't go to anyone else's. If you want to see a super blatant example of this, when Pepsi re-released Crystal Pepsi, Coke released their own "kamikaze" product called Coca-Cola Clear, which they deliberately marketed as a "diet" version specifically to sabotage the very concept of clear cola. They knew it would fail in the market, they only made it to give Pepsi one less product to sell. So yes, businesses do that stuff.