The Gemma series of large models is a series of lightweight large models open-sourced by Google. Just now (March 12, 2025), Google open-sourced the third generation of the Gemma series large models, which includes 4 different parameter scale versions. The third generation, Gemma 3 series, is a multimodal large model, and even the smallest 1 billion parameter scale Gemma 3-1B supports multimodal input.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/de0d1a05-4544-4ae2-ba0f-3c351efc2c52.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/de0d1a05-4544-4ae2-ba0f-3c351efc2c52.png"/>
<ul>
<li>Introduction and features of the Gemma 3 series models</li>
<li>Excellent evaluation of the Gemma 3 series models</li>
<li>Open-source status of Gemma 3</li>
</ul>
<h2>Introduction and features of the Gemma 3 series models</h2>
The Gemma series large models share the same technical origin as Google's Gemini series models but are open-sourced under a free commercial use license. It was first open-sourced in February 2024, with only the 2 billion parameter scale Gemma 2B model and the 7 billion parameter scale Gemma 7B model, and the context length was only 8K. In May 2024, Google open-sourced the Gemma2 series, expanding the range to 3 versions: 2B, 9B, and 27B scales.
Among them, the Gemma 3 - 27B version model was trained on a dataset of 14 trillion, Gemma 3-12B was trained on a dataset of 12 trillion, and the remaining 2 versions were trained on datasets of 4 trillion and 2 trillion, respectively.
Gemma 3's vocabulary expression has reached 262K, indicating very strong representational capability.
Ten months later, Google has open-sourced the third generation of the Gemma 3 series large models, increasing the version range to 4, namely 1B, 4B, 12B, and 27B, and upgrading from a purely large language model to a multimodal large model, which supports image and video input.
This upgrade of Gemma 3 is significant, summarized as follows:
<ul>
<li>The Gemma 3 series models support up to 128K context input (the 1 billion parameter version Gemma 3-1B only supports 32K)</li>
<li>The Gemma 3 series supports more than 140 languages</li>
<li>The Gemma 3 series large models support multimodal input, including text, image, and video input</li>
<li>The Gemma 3 series supports function/tool calls</li>
</ul>
<h2>Excellent evaluation of the Gemma 3 series models</h2>
The Gemma 3 series models include 4 versions, each of which has open-sourced a pre-trained base version (pt suffix version, indicating pre-training) and a version fine-tuned for instructions (it suffix version, indicating instruction fine-tuned), meaning a total of 8 versions of large models have been open-sourced
The largest parameter scale of Gemma 3-27B IT has a fp16 precision size of 54.8GB, and after int8 quantization, it is 27GB, which can be used with two 4090 GPUs. After INT4 quantization, it requires 14GB of video memory, which is completely fine with a single 4090.
The evaluation results of this version of the model are very good, scoring 1338 points on the large model anonymous arena (Chatbot Arena) as of March 8, 2025, ranking 9th globally, only behind the o1-2024-12-17 model, surpassing Qwen2.5-Max and DeepSeek V3.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/1938856e-2265-4f69-91fc-7823758846a9.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/1938856e-2265-4f69-91fc-7823758846a9.png"/>
In other regular evaluations, it also performed well, surpassing Qwen2.5-72B and being very close to DeepSeek V3.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a9815448-2f60-49cd-99cb-9d03548f458f.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a9815448-2f60-49cd-99cb-9d03548f458f.png"/>
According to Google's official statement, the improvements in the Gemma 3 series are significant, with the Gemma 3-4B version model level approaching Gemma 2-27B, while Gemma 3-27B is close to Gemini 1.5-Pro!
<h2>Gemma 3 Open Source Situation</h2>
The 8 models in the Gemma 3 series are open-sourced under the Gemma open-source agreement. Commercial use is allowed, and free licensing is provided.
The ecological adaptation has also been completed, with support from Huggingface, Ollama, Vertex, and llama.cpp.
Author of this article: DataLearner, Source: DataLearner, Original title: &#34;Breaking! Google Open Sources Gemma-3 Model: Supports Multimodal, 128K Input, 27B Version Surpasses DeepSeeK V3 in Large Model Anonymous Arena, Free Commercial License&#34;
Risk Warning and Disclaimer
The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk

GOOGL

GOOG

On March 12, 2025, Google open-sourced the Gemma-3 series of large models, which includes four versions with different parameter scales and supports multimodal input. The Gemma 3-27B version surpassed DeepSeeK V3 in the large model anonymous arena, featuring a 128K context input capability, supporting over 140 languages, and providing free commercial licensing. The training data volume for this series of models reached 14 trillion, demonstrating strong representational capabilities

- Google has released the third generation of the Gemma series of large models, which includes four versions with varying parameter sizes, supporting multimodal inputs.  
- The Gemma 3 series features significant upgrades, such as a maximum context input of 128K and support for over 140 languages, with impressive performance metrics in various evaluations.  
- All eight models in the Gemma 3 series are open-sourced under a commercial-use license, compatible with platforms like Huggingface and Ollama.

Wallstreetcn

Alphabet

Alphabet - C

Vertex

Google open-sources the Gemma-3 model: supports multi-modal, 128K input, 27B version surpasses DeepSeeK V3 in the large model anonymous arena, free commercial use license