# Gemma 3

Gemma 3 is a family of open-weight large language models developed by Google DeepMind, released in March 2025, available in multiple sizes (1B, 4B, 12B, 27B) with multimodal capabilities and long-context support.

Gemma 3 is a family of open-weight [large language models](https://www.wikiprompt.org/wiki/large-language-model) developed by [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind). Released on March 12, 2025, the models are built on the same research and technology as the Gemini 2.0 models, but are designed for developers and researchers to run locally or on their own infrastructure. Gemma 3 includes four primary sizes - 1B, 4B, 12B, and 27B parameters - each available in base and instruction-tuned variants, with a total of 11 variants appearing in public benchmark snapshots as of late 2025.

The models are notable for their multimodal capabilities, accepting both text and image inputs, and for supporting a context window of up to 128,000 tokens. Gemma 3 also includes support for over 140 languages, a significant expansion from earlier Gemma releases. The family is positioned as a competitive open-weight alternative to models from [OpenAI](https://www.wikiprompt.org/wiki/openai) and [Anthropic](https://www.wikiprompt.org/wiki/anthropic), particularly in the sub-30B parameter range.

## Architecture and Training

Gemma 3 models use a [transformer](https://www.wikiprompt.org/wiki/transformer) architecture with [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention), incorporating grouped-query attention for efficiency in larger sizes. The models are trained on a mixture of web documents, code, and mathematical data, with a focus on high-quality filtering. The 27B model uses a SwiGLU activation function and rotary positional embeddings. Training used [Adam](https://www.wikiprompt.org/wiki/adam-optimizer) with a [cosine learning rate schedule](https://www.wikiprompt.org/wiki/learning-rate-schedule) and [gradient clipping](https://www.wikiprompt.org/wiki/gradient-clipping).

Unlike some proprietary models, Gemma 3 weights are released under a permissive license, allowing commercial use and fine-tuning. The instruction-tuned variants were aligned using [RLHF](https://www.wikiprompt.org/wiki/rlaif) (reinforcement learning from human feedback) and supervised fine-tuning, with additional safety filtering.

## Performance and Benchmarks

On public leaderboards, Gemma 3 27B achieves competitive scores on standard benchmarks such as MMLU (massive multitask language understanding), GSM8K (grade school math), and HumanEval (code generation). As of mid-2025, the 27B instruction model outperforms many larger proprietary models on specific reasoning tasks, though it trails the largest frontier models on complex multi-step reasoning. The 4B and 12B sizes are optimized for deployment on edge devices and consumer hardware, with the 1B model suitable for mobile applications.

Gemma 3 also demonstrates strong performance in multilingual tasks, with particular strength in low-resource languages. The models show improved long-context handling compared to predecessors, maintaining coherence over 128K token inputs.

## Deployment and Ecosystem

Gemma 3 is available through multiple channels. It can be accessed via [Google Cloud](https://www.wikiprompt.org/wiki/google-cloud) Vertex AI and [Amazon Web Services](https://www.wikiprompt.org/wiki/amazon-web-services) SageMaker, as well as through [Groq](https://www.wikiprompt.org/wiki/groq) and [SambaNova](https://www.wikiprompt.org/wiki/samba-nova) for high-speed inference. The models are compatible with standard [machine learning](https://www.wikiprompt.org/wiki/machine-learning) frameworks including [Hugging Face Transformers](https://www.wikiprompt.org/wiki/hugging-face) and vLLM.

For local deployment, Gemma 3 supports quantization techniques such as [pruning](https://www.wikiprompt.org/wiki/model-pruning) and [4-bit quantization](https://www.wikiprompt.org/wiki/quantization), enabling the 27B model to run on a single high-end GPU. The smaller variants can run on laptops and Raspberry Pi-class devices. Google provides a Gemma Cookbook with examples for fine-tuning and deployment.

## Safety and Responsible Use

Google DeepMind implemented a layered safety approach for Gemma 3. The instruction-tuned models include safety filters trained to reject harmful prompts, and the company published a model card detailing evaluation on safety benchmarks. However, as an open-weight model, Gemma 3 can be fine-tuned to remove safety guardrails, a trade-off acknowledged by the developers. The license includes a prohibition on using the models for certain high-risk applications, such as weapons development or surveillance.

## Reception and Impact

Gemma 3 received positive reception in the [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) community for its balance of performance and accessibility. Independent evaluations noted that the 27B model often matches or exceeds the quality of larger closed models on specific tasks, while being significantly cheaper to run. The release contributed to the growing trend of open-weight models challenging proprietary systems, alongside efforts from other organizations like [Alibaba](https://www.wikiprompt.org/wiki/alibaba-cloud) and [AI21 Labs](https://www.wikiprompt.org/wiki/ai21-labs).

As of late 2025, Gemma 3 remains a popular choice for research and production applications requiring on-premises or privacy-preserving AI. Its successor, Gemma 3n, was announced in late 2025, but Gemma 3 continues to be widely used in academic and industrial settings.

---
Source: https://www.wikiprompt.org/wiki/gemma-3
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T18:56:46.477012+00:00