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 Gemma series large model, including a total of 4 versions with different parameter scales, the third-generation Gemma 3 series is a multi-modal large model, even the smallest Gemma 3-1B with a scale of 1 billion parameters also supports multi-modal input.

01

Introduction to the Gemma 3 series models and their features

The Gemma series of large models has the same technology as Google’s Gemini series of models, but is open source in the form of a free commercial license. As early as February 2024, it was open-sourced, with only 2 billion parameters for the Gemma 2B model and 7 billion parameters for the Gemma 7B model, and the context length was 8K. In May 2024, Google open-sourced the Gemma2 series, and the range of versions was also increased to 3, which are 2B, 9B, and 27B scales.

Among them, the Gemma 3-27B version model is trained on a 14 trillion dataset, the Gemma 3-12B model is trained on a 12 trillion dataset, and the remaining two versions are 4 trillion data and 2 trillion data training, respectively.

The vocabulary of Gemma 3 is expressed at 262K, which can be said to be very powerful.

Today, 10 months later, Google has open-sourced the third-generation Gemma 3 series of large models, increasing the version range to 4, namely 1B, 4B, 12B and 27B, and upgrading from a pure large language model to a multimodal large model, that is, supporting the input of images and videos.

This time, the upgrade of Gemma 3 is huge, which can be summarized as follows:

Gemma 3 series models support up to 128K contextual input (the 1 billion parameter version of Gemma 3-1B only supports 32K)
The Gemma 3 series supports more than 140 languages
The Gemma 3 series supports multi-modal input, including text, image, and video input
The Gemma 3 series supports function calls/tool calls

02

The Gemma 3 series models are very well reviewed

The Gemma 3 series model consists of 4 versions, each of which is open-sourced with a pre-trained base version (pt suffix version, indicating pre-training) and a fine-tuned instruction version (it suffix version, indicating instruction fine-tuned), that is, a total of 8 versions of the large model have been open-sourced.

The largest parameter size of the Gemma 3-27B IT has an FP16 precision size of 54.8GB, 27GB after int8 quantization, two 4090 cards are available, INT4 requires 14GB of video memory after quantization, and a single 4090 is no problem.

The evaluation results of this version of the model are very good, with a score of 1338 points on the large model Chatbot Arena (as of March 8, 2025), ranking 9th in the world, second only to the o1-2024-12-17 model, surpassing Qwen 2.5-Max and DeepSeek V3.

In other regular tests, the performance is also very good, and the evaluation results surpass the Qwen 2.5-72B, and are very close to DeepSeek V3 and others.

Data source: DataLearnerAI’s large model evaluation and comparison tool

According to Google’s official statement, the Gemma 3 series has been significantly improved, among which the Gemma 3-4B version model level is close to the Gemma 2-27B, and the Gemma 3-27B is close to the Gemini 1.5-Pro!

03

Gemma 3 was severely hallucinated

A small case measured that the Gemma 3 hallucination is a bit serious, not sure if it is too compliant with the instructions. Like asked 😅 Gemma 3 27B, when will llama 4 be released, it says that as of November 21, 2024, Meta has not released llama4, which is very good at this time. But the reply said no, it was released in October 2024, check it out, and Gemma 3-27B replied that I was right 😅😅😅😅

And also nonsense a Llama4 URL!

Gemma 3’s reply to llama4, which has not yet been released

Similar problems, DeepSeek V3 performs much better, I don’t know, I don’t know.

DeepSeek V3’s response to llama4, which has not yet been released

04

Gemma 3 open source

The 8 models of the Gemma 3 series are open source under the Gemma open source license. Commercial use is allowed, and free licensing is allowed.

The ecological adaptation has also been completed long ago, and Huggingface, Ollama, Vertex, and llama.cpp are already supported.

For more information about the Gemma 3 series models and the official website, please refer to DataLearnerAI’s Gemma3 series model information card:

Gemma3-27B：https://www.datalearner.com/ai-models/pretrained-models/gemma-3-27b-it
Gemma3-12B：https://www.datalearner.com/ai-models/pretrained-models/gemma-3-12b-it
Gemma3-4B：https://www.datalearner.com/ai-models/pretrained-models/gemma-3-4b-it
Gemma3-1B：https://www.datalearner.com/ai-models/pretrained-models/gemma-3-1b

Author：Datalearner
Source：Google开源第三代Gemma-3系列模型：支持多模态、最多128K输入，其中Gemma 3-27B在大模型匿名竞技场得分超过了Qwen2.5-Max
The copyright belongs to the author. For commercial reprints, please contact the author for authorization. For non-commercial reprints, please indicate the source.