Google Gemma: The Complete Open-Source AI Model Family

When Google DeepMind released its Gemma model family in February 2024, many observers considered it a modest gesture toward the open-source AI community — a lightweight companion to the proprietary Gemini lineup. Two years and four generations later, that narrative has been turned on its head. The Gemma family has evolved into one of the most powerful, versatile, and widely adopted open-weight model ecosystems in the world, with over 400 million cumulative downloads and more than 100,000 community-built variants on Hugging Face. From edge devices running on a Raspberry Pi to cloud servers crunching frontier-level benchmarks, Gemma now spans a remarkable range of capabilities and specialized use cases.

This article is your definitive guide to the entire Gemma universe — every core generation, every specialized variant, and the architecture decisions that make it all tick.

The Origins: What Is Gemma?

The name "Gemma" comes from the Latin word for precious stone, and like its namesake, the family is built to be both valuable and refined. According to Google DeepMind, Gemma models are purpose-built lightweight AI models that share the same research foundations as Google's proprietary Gemini models, but are packaged as open-weight releases — meaning anyone can download the actual model files, run them locally, and fine-tune them for their own needs.

This distinction matters enormously. Unlike API-only AI models, open-weight models give developers complete control over their data, infrastructure, and deployment environment. They can run entirely on-premises, with no cloud dependency, no per-token billing, and no vendor lock-in. This has made Gemma particularly attractive to enterprises with strict data privacy requirements, academic researchers on limited budgets, and developers building for edge hardware.

Generation by Generation: The Core Gemma Lineage

Gemma 1 (February 2024)

According to Wikipedia, Google launched the original Gemma on February 21, 2024, with just two model sizes: 2B and 7B parameters. At the time, both were text-only and English-focused, but they offered competitive performance for their size and marked a pivotal shift in Google's previously closed approach to AI. The release was widely viewed as a response to Meta's growing open-source AI presence through the Llama family.

While Gemma 1 was capable, it lacked multimodal support and multilingual breadth. Its chief value was in establishing the Gemma brand and proving that Google could produce competitive open-weight models.

Gemma 2 (June–July 2024)

Gemma 2 arrived in June 2024 with a significant capability boost, offering models in 2B, 9B, and 27B parameter sizes. The 9B and 27B models were released first, followed by the 2B variant in July 2024. Gemma 2 introduced architectural improvements that increased both efficiency and quality, and the 27B model began demonstrating near-frontier performance — competing against models significantly larger in size.

Gemma 2 still remained primarily English-focused and text-only, but it raised the community's expectations for what an open-weight model could achieve at a modest parameter count.

Gemma 3 (March 2025)

Gemma 3 marked a genuine generational leap. Launched on March 12, 2025, in 1B, 4B, 12B, and 27B sizes, as described by Google's developer documentation, it introduced several landmark capabilities:

Native multimodality: The 4B, 12B, and 27B models could now process both images and text as input, enabling visual question answering, OCR, chart understanding, and more.
128K token context window: A 16x expansion from prior Gemma models, allowing entire books, multi-image prompts, or complex long-form reasoning in a single pass.
140+ language support: The family shifted from English-centric to truly multilingual, covering over 140 languages out of the box.
Function calling and structured output: Laying the groundwork for agentic AI applications that can interact with APIs and tools programmatically.

Google also released Gemma 3n alongside the main family — a specialized sub-line optimized for consumer devices like phones, laptops, and tablets, with aggressive memory efficiency.

Gemma 4 (April 2026): The Current Flagship

The most recent and most capable generation, Gemma 4 was released on April 2, 2026, under a fully permissive Apache 2.0 license — the most commercially open licensing of any Gemma generation. It comes in four distinct sizes:

Model	Parameters	Architecture	Primary Target
E2B	~2B effective (5.1B total)	Dense	Smartphones, Raspberry Pi
E4B	~4B effective	Dense	Consumer laptops, edge devices
26B A4B	26B (4B active)	Mixture of Experts (MoE)	Consumer GPUs
31B	31B	Dense	Workstations, cloud servers

The "Effective" parameter count in E2B and E4B reflects that while embedding tables inflate the total parameter count, the actual compute during inference is much smaller — a key architectural efficiency for edge deployment.

As reported by tech-insider.org, Gemma 4's 31B model currently ranks #3 among all open models on the Arena AI text leaderboard, while the 26B MoE variant holds the #6 spot — outcompeting models 20 times its size. On the AIME 2026 mathematics benchmark, Gemma 4 scored 89.2%, surpassing Meta's Llama 4 (88.3%), DeepSeek V4 (42.5%), and GPT (37.5%). Similar dominance appears on coding (LiveCodeBench: 80.0%), science reasoning (GPQA Diamond: 84.3%), and agentic task benchmarks.

Gemma 4 features across all sizes include:

Native multimodality (text + images + video), with audio support on E2B and E4B
256K token context on the 31B model
Configurable "thinking mode" for step-by-step reasoning via structured <|think|> output
Native function calling for autonomous agent construction
Hybrid attention combining local sliding-window and global attention
140+ language support

The Specialized Gemma Variants: The "Gemmaverse"

Beyond the core numbered generations, Google has built an entire ecosystem of task-specific Gemma variants. This "Gemmaverse" represents one of the most compelling aspects of the family's design philosophy: a shared, high-quality base that can be fine-tuned into domain specialists.

CodeGemma

Designed for software development workflows, CodeGemma comes in 2B and 7B sizes and is trained on 500 billion tokens of code-heavy data. It supports smart code completion, function generation, multi-language programming (Python, JavaScript, Java, Kotlin, C++, Rust, and more), and can serve as a fully local code assistant — critical for organizations with strict IP or data policies around cloud-based AI coding tools.

PaliGemma and PaliGemma 2

PaliGemma is a vision-language model (VLM) built by combining Google's SigLIP vision encoder with Gemma's language backbone. It is specifically designed to be fine-tuned on custom image datasets, making it highly suitable for specialized visual tasks like medical imaging, satellite image interpretation, document OCR, object detection, and image captioning. PaliGemma 2 built on this foundation with improved performance and multiple resolution options.

MedGemma

MedGemma is Google's healthcare-focused Gemma variant, built on Gemma 3 and trained for performance on medical text and image comprehension. It comes in two configurations: a 4B multimodal version capable of analyzing medical images such as mammograms and X-rays, and a 27B text-only version for clinical text reasoning. Google explicitly notes that MedGemma is not yet clinical-grade and should not replace professional medical judgment — but developers have already used it to build AI-assisted diabetes management apps and radiology report tools.

ShieldGemma and ShieldGemma 2

ShieldGemma serves as a safety classifier for AI pipelines. The original version evaluates the safety of text inputs and outputs against configurable policies, acting as a content moderation layer for other AI applications. ShieldGemma 2 (4B), based on Gemma 3, extends this capability to image safety classification — detecting violent, dangerous, or sexually explicit imagery. It's a key component in Google's Responsible Generative AI Toolkit.

TranslateGemma

Released in January 2026, TranslateGemma is a collection of translation-focused models available in 4B, 12B, and 27B sizes. It is designed to help people communicate across 55 languages with high quality and cultural fidelity, making it a strong candidate for multilingual enterprise applications.

DataGemma

DataGemma is a research-oriented variant that specializes in grounding AI outputs in real-world statistical data. It is designed to reduce hallucinations when dealing with numerical claims by anchoring responses to structured data sources like the Google Data Commons platform.

RecurrentGemma

An architecture-level experiment, RecurrentGemma replaces the standard Transformer attention mechanism with a recurrent neural network (RNN) combined with local attention. The result is a model with significantly lower memory requirements during inference — important for environments where memory is the binding constraint, such as embedded hardware or very long sequence processing at scale.

Gemma Scope

Gemma Scope is not a generative model but an interpretability research suite built on top of Gemma 3. It provides tools for the AI safety research community to study the internal representations of language models — understanding why they behave as they do. This reflects Google's stated commitment to transparent and responsible AI development.

T5Gemma

T5Gemma is a family of encoder-decoder models — a different architecture from the decoder-only design of most Gemma variants. Encoder-decoder models are particularly well-suited for structured tasks like summarization, translation, and question answering where there is a clear input-output mapping. T5Gemma provides a quality-efficiency tradeoff that makes it attractive for production NLP pipelines.

What Makes the Apache 2.0 License a Big Deal

One of the most strategically significant decisions in the Gemma 4 release was moving to the Apache 2.0 open-source license. As explained by MindStudio, Apache 2.0 is one of the most permissive licenses available, allowing developers to use, modify, distribute, and commercially deploy Gemma 4 without royalties, user count caps, or restrictions on derivative works. The only requirement is attribution.

This is a meaningful departure from models like Meta's Llama, which have historically imposed commercial use restrictions above certain active user thresholds. For businesses evaluating open-weight models for production deployment, Gemma 4's licensing clarity removes a significant legal risk factor from the equation.

Ecosystem and Tooling

Part of what makes the Gemma family so practical to work with is its deep integration with the broader AI developer ecosystem. Google's release confirmed day-one support for:

Frameworks: Hugging Face Transformers, TRL, Candle, Keras, PyTorch, JAX
Inference engines: vLLM, llama.cpp, MLX, Ollama, SGLang, LM Studio
Cloud: Google AI Studio, Vertex AI, NVIDIA NIM, Google Kubernetes Engine
Edge: LiteRT-LM, Android ML Kit GenAI Prompt API, AI Edge Gallery, Android Studio Agent Mode
Fine-tuning: Google Colab, Vertex AI, Unsloth, Tunix

This means developers can go from downloading a model to running inference in minutes, on virtually any hardware from a gaming laptop to a data center.

The Competitive Landscape

Gemma 4 arrives in a genuinely crowded open-model market. Meta's Llama 4, released in early 2026, established new performance baselines. China's DeepSeek V4 continues to push the frontier of open-weight efficiency. Alibaba's Qwen 3 and Mistral's latest European models round out the field.

What distinguishes Gemma 4 is its intelligence-per-parameter efficiency. As Stanford's Center for Research on Foundation Models Director Percy Liang observed, as reported by tech-insider.org, architectural innovation and training efficiency can now deliver frontier performance at a fraction of the compute cost — and Gemma 4's edge models outperforming Gemma 3's 27B model while being 6–10x smaller is the clearest demonstration of that trend.

Looking Ahead

The trajectory of the Gemma family tells a clear story: open-weight models are closing the gap with proprietary systems faster than anyone expected. What began as a modest 2B-parameter text model in early 2024 has grown, within two years, into a multi-generational ecosystem with specialized variants for healthcare, safety, coding, translation, and scientific research — all under fully permissive open-source licensing.

For developers, researchers, and enterprises, the Gemma family now represents one of the most complete, practically deployable open-source AI stacks available anywhere. Whether you need a 2-billion-parameter model running silently on a smartphone or a 31-billion-parameter powerhouse challenging proprietary frontier models, there is a Gemma variant built for the job.

The "Gemmaverse" is not just growing — it is becoming a serious alternative to the closed AI ecosystem entirely.

google gemma

Related Topics

Google Gemma: The Complete Open-Source AI Model Family