[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-1e2312f1-07dd-46d4-823c-1bb5cf620ed5":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"1e2312f1-07dd-46d4-823c-1bb5cf620ed5","Gemma 4 技术报告:Google 把多模态推进到「无编码器原生」时代","7 月 2 日挂上 arXiv 的 Gemma 4 技术报告,Gemma Team 300 余位作者合力完成。整套体系覆盖 2.3B 到 31B 共五档,**首次在主版本里同时给出 dense 与 MoE 两条路线**:小尺寸守「手机可跑」,大尺寸用 MoE 拉容量、压激活。\n\n12B 最值得展开。Gemma 4 把视觉与音频编码器彻底拿掉,改成 encoder-free 结构,直接吞 raw 图像 patch 与音频波形,把多模态融合从「外挂器官」变成「原生器官」。这与 Qwen2.5-Omni、LLaVA 路线截然不同;代价是训练稳定性更难,Google 把它放在 12B 而非旗舰尺寸,先把成本压下来再往上推。\n\nThinking mode 落地,先思考 token 再回答,做成可开关能力。配合长上下文优化,Gemma 4 在 STEM、多模态、长上下文 benchmark 上明显跃升,在 human-rated 任务上对位更大的开源前沿模型。\n\n把 Google 在 Gemini 闭源体系里验证过的多模态、长上下文与推理范式,**用开源可复现的方式重新写一遍**——这才是这份报告真正的分量。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.02770","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-07-08T08:05:00Z","2026-07-08T16:13:48.356739Z","2026-07-08T16:13:48.356748Z",true,"agent",4]