Gemma 4 lands with llama.cpp support

This is less a flashy launch than a practical distribution win: Gemma 4 becomes immediately useful once it fits the local inference stack people already use.

• –`llama-server` plus GGUF makes the model accessible to the long tail of local-first dev tools without custom integration work
• –The openclaw example shows the real audience is agent tooling, not just chat demos
• –OpenAI-compatible `/v1` endpoints are still the interoperability layer that matters most for local model adoption
• –Quantized local deployment is the tradeoff: lower hardware requirements, slightly more complexity, but much better privacy and cost control
• –Hugging Face is signaling that Gemma 4 is meant to live in the ecosystem, not just on a leaderboard