The Gemma 4 tech report, posted to arXiv on July 2, was completed by over 300 authors from the Gemma Team. The system covers five scales from 2.3B to 31B, for the first time in the main version simultaneously providing both dense and MoE paths: the small sizes keep "runnable on phones", while the large sizes use MoE to expand capacity and compress activation. The 12B is most worth elaborating. Gemma 4 completely removes the visual and audio encoders, switching to an encoder-free structure, directly swallowing raw image patches and audio waveforms, making multimodal fusion from "external organ" to "native organ". This is completely different from the Qwen2.5-Omni, LLaVA route; the cost is harder training stability, and Google puts it at 12B rather than the flagship size, first getting the cost down before pushing it up. Thinking mode lands, first think tokens then answer, made into a switchable capability. Combined with long-context optimization, Gemma 4 makes a clear jump on STEM, multimodal, and long-context benchmarks, and on human-rated tasks it stands opposite larger open-source frontier models. The point of taking the multimodal, long-context, and reasoning paradigms validated by Google in the closed-source Gemini system and rewriting them in an open-source reproducible way — that's the real weight of this report.