[[["容易理解","easyToUnderstand","thumb-up"],["確實解決了我的問題","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["缺少我需要的資訊","missingTheInformationINeed","thumb-down"],["過於複雜/步驟過多","tooComplicatedTooManySteps","thumb-down"],["過時","outOfDate","thumb-down"],["翻譯問題","translationIssue","thumb-down"],["示例/程式碼問題","samplesCodeIssue","thumb-down"],["其他","otherDown","thumb-down"]],["上次更新時間:2025-03-21 (世界標準時間)。"],[],[],null,["# Gemma 3 model overview\n\nGemma is a family of generative artificial intelligence (AI) models and you can\nuse them in a wide variety of generation tasks, including question answering,\nsummarization, and reasoning. Gemma models are provided with open weights and\npermit responsible\n[commercial use](/gemma/terms),\nallowing you to tune and deploy them in your own projects and applications.\n\nThe Gemma 3 release includes the following key features. Try it in\n[AI Studio](https://aistudio.google.com/prompts/new_chat?model=gemma-3-27b-it):\n\n- [**Image and text input**](#multimodal-input): Multimodal capabilities let you input images and text to understand and analyze visual data. [Start building](/gemma/docs/core/keras_inference)\n- [**128K token context**](#128k-context): Significantly large input context for analyzing more data and solving more complex problems.\n- [**Function calling**](#function-calling): Build natural language interfaces for working with programming interfaces. [Start building](/gemma/docs/capabilities/function-calling)\n- [**Wide language support**](#multilingual): Work in your language or expand your AI application's language capabilities with support for over 140 languages. [Start building](/gemma/docs/spoken-language)\n- [**Developer friendly model sizes**](#sizes): Choose a model size (270M, 1B, 4B, 12B, 27B) and precision level that works best for your task and compute resources.\n\nYou can download Gemma 3 models from\n[Kaggle](https://www.kaggle.com/models?query=gemma3&publisher=google) and\n[Hugging Face](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d).\nFor more technical details on Gemma 3, see the\n[Model Card](/gemma/docs/core/model_card_3) and\n[Technical Report](https://goo.gle/Gemma3Report).\nEarlier versions of Gemma core models are also available for download. For more\ninformation, see [Previous Gemma models](#previous-models).\n\n[Try Gemma 3](https://aistudio.google.com/prompts/new_chat?model=gemma-3-27b-it)\n[Get it on Kaggle](https://www.kaggle.com/models?query=gemma3&publisher=google)\n[Get it on Hugging Face](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d)\n\nMultimodal image and text input\n-------------------------------\n\nYou can tackle complex analysis and generation tasks with Gemma 3 with its\nability to handle image and text data. You can use the model to interpret image\ndata, identify objects, extract text data, and complete many other visual input\nto text output tasks.\n[Start building](/gemma/docs/core/keras_inference)\n| **Important:** The Gemma 3 270M and 1B models are text only and *do not support\n| image input*.\n\n128K token context window\n-------------------------\n\nGemma 3 models (4B, 12B, and 27B) can handle prompt inputs up to 128K tokens, a\n16x larger context window than previous Gemma models. The large number of tokens\nmeans you can process several, multi page articles, larger single articles, or\nhundreds of images in a single prompt.\n| **Important:** The Gemma 3 270M and 1B models can process up to 32k tokens.\n\nWide language support\n---------------------\n\nWork in your own language with built-in support for over 140 languages. Gemma 3\nis trained to support a large number of languages compared to previous Gemma\nversions, letting you take on more visual and text tasks in the languages your\ncustomers use.\n[Start building](/gemma/docs/spoken-language)\n\nFunction calling\n----------------\n\nBuild intelligent, natural language controls for programming interfaces. Gemma\n3 lets you define coding functions with specific syntax and constraints, and\nthe model can call these functions to complete tasks.\n[Start building](/gemma/docs/capabilities/function-calling)\n\nParameter sizes and quantization\n--------------------------------\n\nGemma 3 models are available in 5 parameter sizes: 270M, 1B, 4B, 12B, and 27B.\nThe models can be used with their default precision (16-bit) or with a lower\nprecision using quantization. The different sizes and precisions represent a set\nof trade-offs for your AI application. Models with higher parameters and bit\ncounts (higher precision) are generally more capable, but are more expensive to\nrun in terms of processing cycles, memory cost and power consumption. Models\nwith lower parameters and bit counts (lower precision) have less capabilities,\nbut may be sufficient for your AI task.\n\nFor all Gemma 3 models, [Quantization-Aware Trained](https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/)\ncheckpoints are provided, which allow quantizing (reducing the precision), while\npreserving high-quality.\n\nThe following table details the approximate GPU or TPU memory requirements for\nrunning inference with each size of the Gemma 3 model versions. Note that the\nnumbers may changed based on inference tool.\n\n| **Parameters** | **BF16 (16-bit)** | **SFP8 (8-bit)** | **Q4_0 (4-bit)** |\n|----------------------------|-------------------|------------------|------------------|\n| Gemma 3 270M (*text only*) | 400 MB | 297 MB | 240 MB |\n| Gemma 3 1B (*text only*) | 1.5 GB | 1.1 GB | 892 MB |\n| Gemma 3 4B | 6.4 GB | 4.4 GB | 3.4 GB |\n| Gemma 3 12B | 20 GB | 12.2 GB | 8.7 GB |\n| Gemma 3 27B | 46.4 GB | 29.1 GB | 21 GB |\n\n**Table 1.** Approximate GPU or TPU memory required to load Gemma 3 models\nbased on parameter count and quantization level.\n| **Caution:** These estimates only include the memory required to load the models. They don't include the additional memory required for the prompt tokens or supporting software.\n\nMemory consumption increases based on the total number of tokens required for\nthe prompt you run. The larger the number of tokens required to process your\nprompt, the higher the memory required, which is in addition to the memory\nrequired to load the model.\n| **Note:** Memory requirements for *fine-tuning* Gemma models are significantly higher than running inference. The requirements depend on the development framework and tuning technique you use, such as Low Rank Adapter (LoRA) versus full-precision tuning.\n\nPrevious Gemma models\n---------------------\n\nYou can work with previous generations of Gemma models, which are also\navailable from [Kaggle](https://www.kaggle.com/models?query=gemma) and\n[Hugging Face](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d).\nFor more technical details about previous Gemma models, see the following\nmodel card pages:\n\n- Gemma 2 [Model Card](/gemma/docs/core/model_card_2)\n- Gemma 1 [Model Card](/gemma/docs/core/model_card)\n\nReady to start building?\n[Get started](/gemma/docs/get_started)\nwith Gemma models!"]]