{"id":11486,"date":"2026-04-05T18:18:55","date_gmt":"2026-04-05T09:18:55","guid":{"rendered":"https:\/\/www.stechstar.com\/user\/wordpress\/?p=11486"},"modified":"2026-09-14T14:48:05","modified_gmt":"2026-09-14T05:48:05","slug":"%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ea%b8%b0%ec%88%a0-google-%ec%98%a4%ed%94%88-%eb%aa%a8%eb%8d%b8-gemma-4-%ea%b3%b5%ea%b0%9c","status":"publish","type":"post","link":"https:\/\/www.stechstar.com\/user\/wordpress\/%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ea%b8%b0%ec%88%a0-google-%ec%98%a4%ed%94%88-%eb%aa%a8%eb%8d%b8-gemma-4-%ea%b3%b5%ea%b0%9c\/","title":{"rendered":"[\uc778\uacf5\uc9c0\ub2a5 \uae30\uc220] Google, \uc624\ud508 \ubaa8\ub378 Gemma 4 \uacf5\uac1c\u00a0"},"content":{"rendered":"

[\uc778\uacf5\uc9c0\ub2a5 \uae30\uc220] Google, \uc624\ud508 \ubaa8\ub378 Gemma 4 \uacf5\uac1c <\/h1>\n
\n

Google, \uc624\ud508 \ubaa8\ub378 Gemma 4 \uacf5\uac1c<\/a><\/h1>\n

 (deepmind.google)<\/span><\/div>\n

8<\/span>P by GN\u207a<\/a> 2\uc77c\uc804<\/span> | \u2605 favorite<\/a><\/span> | \ub313\uae00 2\uac1c<\/a><\/div>\n
\n
\n
\n
    \n
  • Google DeepMind\uac00 Gemini 3 \uae30\uc220\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \ucc28\uc138\ub300 \uc624\ud508 AI \ubaa8\ub378 Gemma 4<\/strong>\ub97c \ubc1c\ud45c, \ub9e4\uac1c\ubcc0\uc218\ub2f9 \uc9c0\ub2a5 \ud6a8\uc728<\/strong>\uc744 \uadf9\ub300\ud654\ud55c \uad6c\uc870\ub85c \uc124\uacc4\ub428<\/li>\n
  • \ubaa8\ub378\uc740 E2B, E4B, 26B, 31B<\/strong> \ub124 \uac00\uc9c0 \ud06c\uae30\ub85c \uc81c\uacf5\ub418\uba70, \ubaa8\ubc14\uc77c\u00b7IoT\ubd80\ud130 \uac1c\uc778\uc6a9 GPU \ud658\uacbd\uae4c\uc9c0<\/strong> \ud3ed\ub113\uc740 \uc2e4\ud589 \ubc94\uc704\ub97c \uc9c0\uc6d0<\/li>\n
  • \uba40\ud2f0\ubaa8\ub2ec \ucd94\ub860, 140\uac1c \uc5b8\uc5b4 \uc9c0\uc6d0, \uc5d0\uc774\uc804\ud2b8\ud615 \uc6cc\ud06c\ud50c\ub85c, \uc138\ubc00\ud55c \ud30c\uc778\ud29c\ub2dd, \ud6a8\uc728\uc801 \uc544\ud0a4\ud14d\ucc98<\/strong> \ub4f1 \uc8fc\uc694 \uae30\ub2a5\uc744 \ud3ec\ud568<\/li>\n
  • \uc218\ud559\u00b7\ucf54\ub529\u00b7\uba40\ud2f0\ubaa8\ub2ec \uc774\ud574<\/strong> \uc601\uc5ed\uc5d0\uc11c Gemma 3 \ub300\ube44 \uc131\ub2a5\uc774 \ud06c\uac8c \ud5a5\uc0c1\ub418\uc5c8\uc73c\uba70, \ubcf4\uc548\u00b7\uc2e0\ub8b0\uc131 \uae30\uc900<\/strong>\uc740 Google \uc0c1\uc6a9 \ubaa8\ub378\uacfc \ub3d9\uc77c \uc218\uc900 \uc720\uc9c0<\/li>\n
  • \ubaa8\ub378 \uac00\uc911\uce58\ub294 Hugging Face, Ollama, Kaggle, LM Studio, Docker<\/strong> \ub4f1\uc5d0\uc11c \ub2e4\uc6b4\ub85c\ub4dc \uac00\ub2a5\ud558\uba70, \ub85c\uceec \ubc0f \ud074\ub77c\uc6b0\ub4dc \ud658\uacbd \ud1b5\ud569 \uc2e4\ud589<\/strong>\uc744 \uc9c0\uc6d0\ud568<\/li>\n<\/ul>\n
    \n

    Gemma 4 \u2014 \ucc28\uc138\ub300 \uc624\ud508 AI \ubaa8\ub378<\/h2>\n
      \n
    • Gemma 4<\/strong>\ub294 Gemini 3\uc758 \uc5f0\uad6c\uc640 \uae30\uc220<\/strong>\uc744 \uae30\ubc18\uc73c\ub85c \uac1c\ubc1c\ub41c Google DeepMind\uc758 \ucd5c\uc2e0 \uc624\ud508 \ubaa8\ub378\ub85c, \ub9e4\uac1c\ubcc0\uc218\ub2f9 \uc9c0\ub2a5 \ud6a8\uc728(intelligence-per-parameter)<\/strong> \uc744 \uadf9\ub300\ud654\ud55c \uad6c\uc870\ub97c \uac00\uc9d0<\/li>\n
    • \ubaa8\ub378\uc740 E2B, E4B, 26B, 31B<\/strong> \ub124 \uac00\uc9c0 \ud06c\uae30\ub85c \uc81c\uacf5\ub418\uba70, \ubaa8\ubc14\uc77c\u00b7IoT\ubd80\ud130 \uac1c\uc778\uc6a9 \uc6cc\ud06c\uc2a4\ud14c\uc774\uc158\uae4c\uc9c0 \ub2e4\uc591\ud55c \ud658\uacbd\uc5d0\uc11c \uc2e4\ud589 \uac00\ub2a5<\/li>\n
    • \uba40\ud2f0\ubaa8\ub2ec \ucd94\ub860<\/strong>, 140\uac1c \uc5b8\uc5b4 \uc9c0\uc6d0<\/strong>, \uc5d0\uc774\uc804\ud2b8\ud615 \uc6cc\ud06c\ud50c\ub85c<\/strong>, \uc138\ubc00\ud55c \ud30c\uc778\ud29c\ub2dd<\/strong>, \ud6a8\uc728\uc801 \uc544\ud0a4\ud14d\ucc98<\/strong>\ub97c \uc8fc\uc694 \uae30\ub2a5\uc73c\ub85c \ud3ec\ud568<\/li>\n
    • \uc131\ub2a5 \ubca4\uce58\ub9c8\ud06c<\/strong>\uc5d0\uc11c Gemma 3 \ub300\ube44 \uc804\ubc18\uc801\uc778 \ud5a5\uc0c1\uce58\ub97c \uae30\ub85d\ud558\uba70, \ud2b9\ud788 \uc218\ud559\u00b7\ucf54\ub529\u00b7\uba40\ud2f0\ubaa8\ub2ec \uc774\ud574 \uc601\uc5ed\uc5d0\uc11c \ub192\uc740 \uc810\uc218\ub97c \ub2ec\uc131<\/li>\n
    • \ubcf4\uc548\u00b7\uc2e0\ub8b0\uc131 \uae30\uc900<\/strong>\uc740 Google\uc758 \uc0c1\uc6a9 \ubaa8\ub378\uacfc \ub3d9\uc77c \uc218\uc900\uc73c\ub85c \uc720\uc9c0\ub418\uba70, Hugging Face, Ollama, Kaggle, LM Studio, Docker \ub4f1\uc5d0\uc11c \ubaa8\ub378 \uac00\uc911\uce58\ub97c \ub2e4\uc6b4\ub85c\ub4dc \uac00\ub2a5<\/li>\n<\/ul>\n

      \ubaa8\ub378 \uad6c\uc131 \ubc0f \ud6a8\uc728\uc131<\/h2>\n
        \n
      • Gemma 4\ub294 Gemini 3\uc758 \uae30\uc220 \uae30\ubc18<\/strong>\uc73c\ub85c \uc124\uacc4\ub418\uc5b4 \uc9c0\ub2a5 \ud6a8\uc728\uc744 \uadf9\ub300\ud654<\/strong>\ud55c \uc624\ud508 \ubaa8\ub378 \uad6c\uc870\ub97c \ucc44\ud0dd<\/li>\n
      • \ubaa8\ub378 \ud06c\uae30\ub294 E2B, E4B, 26B, 31B<\/strong> \ub124 \uac00\uc9c0 \ubc84\uc804\uc73c\ub85c \uad6c\ubd84\ub418\uba70, \uac01 \ubc84\uc804\uc740 \ucef4\ud4e8\ud305 \uc790\uc6d0\uacfc \uba54\ubaa8\ub9ac \ud6a8\uc728\uc131<\/strong>\uc5d0 \ub530\ub77c \ucd5c\uc801\ud654\ub428\n
          \n
        • E2B\u00b7E4B<\/strong>: \ubaa8\ubc14\uc77c \ubc0f IoT \uae30\uae30\uc6a9\uc73c\ub85c, \ucd5c\ub300 \ud6a8\uc728\uc131\uacfc \uc624\ud504\ub77c\uc778 \uc2e4\ud589<\/strong> \uc9c0\uc6d0<\/li>\n
        • 26B\u00b731B<\/strong>: \uac1c\uc778\uc6a9 GPU \ud658\uacbd\uc5d0\uc11c \ud504\ub860\ud2f0\uc5b4\uae09 \ucd94\ub860 \ub2a5\ub825<\/strong> \uc81c\uacf5<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

          \uc8fc\uc694 \uae30\ub2a5<\/h2>\n
            \n
          • \n

            Agentic workflows<\/h3>\n
              \n
            • \ud568\uc218 \ud638\ucd9c(function calling)<\/strong> \uc744 \ub124\uc774\ud2f0\ube0c\ub85c \uc9c0\uc6d0\ud574, \uc0ac\uc6a9\uc790\ub97c \ub300\uc2e0\ud574 \uacc4\ud68d\u00b7\uc571 \ud0d0\uc0c9\u00b7\uc791\uc5c5 \uc218\ud589<\/strong>\uc774 \uac00\ub2a5\ud55c \uc790\uc728\ud615 \uc5d0\uc774\uc804\ud2b8 \uad6c\ucd95 \uac00\ub2a5<\/li>\n<\/ul>\n<\/li>\n
            • \n

              Multimodal reasoning<\/h3>\n
                \n
              • \uc624\ub514\uc624\uc640 \ube44\uc8fc\uc5bc \uc774\ud574 \ub2a5\ub825<\/strong>\uc744 \uacb0\ud569\ud574 \ud48d\ubd80\ud55c \uba40\ud2f0\ubaa8\ub2ec \uc560\ud50c\ub9ac\ucf00\uc774\uc158 \uac1c\ubc1c<\/strong> \uc9c0\uc6d0<\/li>\n<\/ul>\n<\/li>\n
              • \n

                Support for 140 languages<\/h3>\n
                  \n
                • \ub2e8\uc21c \ubc88\uc5ed\uc744 \ub118\uc5b4 \ubb38\ud654\uc801 \ub9e5\ub77d \uc774\ud574<\/strong>\ub97c \ud3ec\ud568\ud55c \ub2e4\uad6d\uc5b4 \uacbd\ud5d8 \uc0dd\uc131 \uac00\ub2a5<\/li>\n<\/ul>\n<\/li>\n
                • \n

                  Fine tuning<\/h3>\n
                    \n
                  • \uc0ac\uc6a9\uc790\uac00 \uc120\ud638\ud558\ub294 \ud504\ub808\uc784\uc6cc\ud06c\uc640 \uae30\ubc95\uc73c\ub85c \ud2b9\uc815 \uc791\uc5c5 \uc131\ub2a5 \ud5a5\uc0c1<\/strong>\uc744 \uc704\ud55c \ud30c\uc778\ud29c\ub2dd \uac00\ub2a5<\/li>\n<\/ul>\n<\/li>\n
                  • \n

                    Efficient architecture<\/h3>\n
                      \n
                    • \uc790\uccb4 \ud558\ub4dc\uc6e8\uc5b4\uc5d0\uc11c \uc2e4\ud589 \uac00\ub2a5<\/strong>\ud558\uba70, \ud6a8\uc728\uc801\uc778 \uac1c\ubc1c \ubc0f \ubc30\ud3ec \ud658\uacbd \uc81c\uacf5<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

                      \uc131\ub2a5<\/h2>\n
                        \n
                      • Gemma 4\ub294 \ub2e4\uc591\ud55c \ud14d\uc2a4\ud2b8 \uc0dd\uc131 \uad00\ub828 \ub370\uc774\ud130\uc14b\uacfc \uc9c0\ud45c<\/strong>\ub97c \uae30\ubc18\uc73c\ub85c \ud3c9\uac00\ub428<\/li>\n
                      • \uc8fc\uc694 \ubca4\uce58\ub9c8\ud06c \uacb0\uacfc (Gemma 4 31B IT \uae30\uc900):\n
                          \n
                        • Arena AI (text)<\/strong>: 1452 (Gemma 3 27B \ub300\ube44 1365)<\/li>\n
                        • MMMLU (\ub2e4\uad6d\uc5b4 Q&A)<\/strong>: 85.2%<\/li>\n
                        • MMMU Pro (\uba40\ud2f0\ubaa8\ub2ec \ucd94\ub860)<\/strong>: 76.9%<\/li>\n
                        • AIME 2026 (\uc218\ud559)<\/strong>: 89.2%<\/li>\n
                        • LiveCodeBench v6 (\ucf54\ub529 \ubb38\uc81c)<\/strong>: 80.0%<\/li>\n
                        • GPQA Diamond (\uacfc\ud559 \uc9c0\uc2dd)<\/strong>: 84.3%<\/li>\n
                        • \u03c42-bench (\uc5d0\uc774\uc804\ud2b8 \ub3c4\uad6c \uc0ac\uc6a9)<\/strong>: 86.4%<\/li>\n<\/ul>\n<\/li>\n
                        • \uc804\ubc18\uc801\uc73c\ub85c Gemma 3 \ub300\ube44 \ubaa8\ub4e0 \ud56d\ubaa9\uc5d0\uc11c \uc131\ub2a5 \ud5a5\uc0c1<\/strong>\uc744 \ubcf4\uc774\uba70, \ud2b9\ud788 \uc218\ud559\u00b7\ucf54\ub529\u00b7\uba40\ud2f0\ubaa8\ub2ec \uc774\ud574<\/strong> \uc601\uc5ed\uc5d0\uc11c \ud070 \uac1c\uc120<\/li>\n<\/ul>\n

                          E2B \ubc0f E4B \u2014 \ubaa8\ubc14\uc77c \ubc0f IoT\uc6a9<\/h2>\n
                            \n
                          • \uc624\ub514\uc624\u00b7\ube44\uc804 \uc9c0\uc6d0<\/strong>\uc744 \ud1b5\ud574 \uc5e3\uc9c0 \ub514\ubc14\uc774\uc2a4\uc5d0\uc11c \uc2e4\uc2dc\uac04 \ucc98\ub9ac<\/strong> \uac00\ub2a5<\/li>\n
                          • \uc2a4\ub9c8\ud2b8\ud3f0, Raspberry Pi, Jetson Nano<\/strong> \ub4f1\uc5d0\uc11c \uc644\uc804 \uc624\ud504\ub77c\uc778 \uc2e4\ud589<\/strong> \ubc0f \uac70\uc758 \uc81c\ub85c \uc9c0\uc5f0(latency)<\/strong> \uc131\ub2a5 \uc81c\uacf5<\/li>\n
                          • Google AI Edge Gallery\ub97c \ud1b5\ud574 \uccb4\ud5d8 \uac00\ub2a5<\/li>\n<\/ul>\n

                            26B \ubc0f 31B \u2014 \uace0\uc131\ub2a5 \ub85c\uceec AI<\/h2>\n
                              \n
                            • IDE, \ucf54\ub529 \uc5b4\uc2dc\uc2a4\ud134\ud2b8, \uc5d0\uc774\uc804\ud2b8\ud615 \uc6cc\ud06c\ud50c\ub85c<\/strong>\uc5d0 \uc801\ud569\ud55c \uace0\uae09 \ucd94\ub860 \uae30\ub2a5 \uc81c\uacf5<\/li>\n
                            • \uc18c\ube44\uc790\uc6a9 GPU\uc5d0 \ucd5c\uc801\ud654<\/strong>\ub418\uc5b4 \ud559\uc0dd\u00b7\uc5f0\uad6c\uc790\u00b7\uac1c\ubc1c\uc790\uac00 \ub85c\uceec AI \uc11c\ubc84 \ud658\uacbd<\/strong>\uc744 \uad6c\ucd95 \uac00\ub2a5<\/li>\n
                            • Google AI Studio\uc5d0\uc11c \uc9c1\uc811 \uc2e4\ud589 \uac00\ub2a5<\/li>\n<\/ul>\n

                              \ubcf4\uc548 \ubc0f \uc2e0\ub8b0\uc131<\/h2>\n
                                \n
                              • Gemma 4\ub294 Google\uc758 \uc0c1\uc6a9 \ubaa8\ub378\uacfc \ub3d9\uc77c\ud55c \uc778\ud504\ub77c \ubcf4\uc548 \ud504\ub85c\ud1a0\ucf5c<\/strong>\uc744 \uc801\uc6a9<\/li>\n
                              • \uae30\uc5c5 \ubc0f \uacf5\uacf5\uae30\uad00<\/strong>\uc774 \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 \ud22c\uba85\ud558\uace0 \uc2e0\ub8b0\ud560 \uc218 \uc788\ub294 \uae30\ubc18<\/strong> \uc81c\uacf5<\/li>\n
                              • \ucd5c\uace0 \uc218\uc900\uc758 \ubcf4\uc548\u00b7\uc2e0\ub8b0\uc131 \uae30\uc900<\/strong>\uc744 \ucda9\uc871\ud558\uba74\uc11c\ub3c4 \ucd5c\uc2e0 AI \uae30\ub2a5<\/strong>\uc744 \uc81c\uacf5<\/li>\n<\/ul>\n

                                \ub2e4\uc6b4\ub85c\ub4dc \ubc0f \uc2e4\ud589<\/h2>\n
                                  \n
                                • \n

                                  \ubaa8\ub378 \uac00\uc911\uce58 \ub2e4\uc6b4\ub85c\ub4dc<\/h3>\n
                                    \n
                                  • Hugging Face<\/strong>, Ollama<\/strong>, Kaggle<\/strong>, LM Studio<\/strong>, Docker Hub<\/strong>\uc5d0\uc11c Gemma 4 \ubaa8\ub378 \uac00\uc911\uce58 \uc81c\uacf5<\/li>\n<\/ul>\n<\/li>\n
                                  • \n

                                    \ud559\uc2b5 \ubc0f \ubc30\ud3ec \uc9c0\uc6d0<\/h3>\n
                                      \n
                                    • Jax<\/strong>, Vertex AI<\/strong>, Keras<\/strong>, Google AI Edge<\/strong>, Google Kubernetes Engine<\/strong>, Ollama<\/strong> \ub4f1 \ub2e4\uc591\ud55c \ud50c\ub7ab\ud3fc\uacfc \ud1b5\ud569 \uc9c0\uc6d0<\/li>\n
                                    • \uacf5\uc2dd \ubb38\uc11c \ubc0f API\ub97c \ud1b5\ud574 \ud6c8\ub828\u00b7\ubc30\ud3ec\u00b7\ucd94\ub860 \ud658\uacbd \uad6c\uc131<\/strong> \uac00\ub2a5<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

                                      Gemmaverse \ucee4\ubba4\ub2c8\ud2f0<\/h2>\n
                                        \n
                                      • Gemmaverse<\/strong>\ub97c \ud1b5\ud574 \uc804 \uc138\uacc4 \uac1c\ubc1c\uc790\ub4e4\uc774 Gemma\ub97c \ud65c\uc6a9\ud574 \uad6c\ucd95\ud55c \ud504\ub85c\uc81d\ud2b8\ub97c \ud0d0\uc0c9 \uac00\ub2a5<\/li>\n
                                      • Google DeepMind\uc758 X, Instagram, YouTube, LinkedIn, GitHub<\/strong> \ucc44\ub110\uc744 \ud1b5\ud574 \ucd5c\uc2e0 \uc5c5\ub370\uc774\ud2b8 \uc81c\uacf5<\/li>\n
                                      • \uad6c\ub3c5\uc744 \ud1b5\ud574 \ucd5c\uc2e0 AI \ud601\uc2e0 \uc18c\uc2dd<\/strong> \uc218\uc2e0 \uac00\ub2a5<\/li>\n<\/ul>\n<\/div>\n<\/div>\n<\/div>\n

                                        [\ucd9c\ucc98] https:\/\/news.hada.io\/topic?id=28138