Gemma 4 12B ships encoder-free multimodal at 12B parameters and trades blows with models twice its size on community benchmarks. The encoder-free architecture is the part worth sitting with. Removing the vision encoder from the pipeline cuts inference overhead and simplifies deployment, and Google doing this at 12B suggests the design is mature enough for production, not just a research demo. The catch: benchmark comparisons against Qwen 2.5 9B show Gemma 4 losing on five of eight tasks, wh...