Gemma, often called Google Gemma, is a family of lightweight open-weight language models developed by Google DeepMind from the same research and technology used to build the Gemini models. Among AI models & assistants, Gemma occupies the space of downloadable models that developers run themselves, on a workstation, a single accelerator, a phone or a private server, rather than a hosted chat product.
Since the first release in February 2024, successive generations have added larger context windows, image understanding and many specialized variants. The weights are published under the Gemma Terms of Use, which permit commercial use but impose a prohibited-use policy, so Gemma is better described as open-weight than open source.
How Gemma models are built
Gemma models are decoder-only transformers that predict the next token, trained on large text corpora and, in later generations, on images paired with text. They share a SentencePiece tokenizer with a vocabulary of roughly 256,000 entries with Gemini, which gives them good coverage of many languages and of code.
Several design choices target efficiency at small sizes. Gemma 2 alternated local sliding-window attention with global attention layers to cut memory use, and its smaller models were trained with knowledge distillation from a larger teacher model rather than from raw next-token targets alone. Gemma 3 raised the ratio of local to global layers to about five to one, which keeps the key-value cache manageable at long context.
Each generation ships in two forms: pretrained base checkpoints for further training and instruction-tuned checkpoints refined with supervised fine-tuning and reinforcement learning for chat and task following.
Gemma generations and model sizes
The Gemma family has grown through several generations, each widening the range of sizes and capabilities. The table summarizes the core releases as published at launch; parameter counts are approximate.
| Generation | Released | Typical sizes | Notable features |
|---|---|---|---|
| Gemma 1 | February 2024 | 2B and 7B | Text only, 8K context |
| Gemma 2 | June 2024 | 2B, 9B and 27B | Distillation, alternating attention |
| Gemma 3 | March 2025 | 1B, 4B, 12B and 27B | Image input from 4B up, up to 128K context |
| Gemma 3n | 2025 | Effective 2B and 4B | Designed for phones, audio and image input |
Gemma 3 added a SigLIP-based vision encoder to its larger models, enabling image description, chart reading and visual question answering. Gemma 3n introduced techniques such as per-layer embeddings and a nested MatFormer architecture so that a model with more raw parameters can run with the memory footprint of a smaller one.
Specialized Gemma variants
Beyond the core models, Google has released purpose-built variants that start from Gemma weights and are tuned or extended for a narrow job. These variants illustrate how an open-weight base can serve as a foundation for domain models.
| Variant | Purpose | Typical use |
|---|---|---|
| CodeGemma | Code completion and generation | IDE assistants, fill-in-the-middle |
| PaliGemma | Vision-language tasks | Captioning, detection, document reading |
| RecurrentGemma | Recurrent architecture for long inputs | Lower memory at long sequence length |
| ShieldGemma | Safety classification | Filtering prompts and responses |
| MedGemma | Medical text and images | Research on clinical assistants |
| EmbeddingGemma | Text embeddings | On-device search and retrieval |
Independent developers have also published thousands of community fine-tunes and quantized builds, which is a direct consequence of the weights being freely downloadable.
How Gemma compares with related model families
Gemma is best understood against Gemini on one side and other open-weight families on the other. Gemini is Google’s flagship, accessible only through hosted services, while Gemma is smaller, self-hostable and less capable on the hardest reasoning tasks.
| Family | Access model | Size focus | Main difference from Gemma |
|---|---|---|---|
| Gemini | Hosted API and apps | Frontier scale | Far more capable, not downloadable |
| Llama | Open weights, custom license | Small to very large | Broader size range, different license terms |
| Phi | Open weights, permissive license | Small models | Emphasis on curated synthetic data |
| Qwen | Open weights, mostly permissive | Small to very large | Strong multilingual and coding focus |
| Mistral | Mixed open and hosted | Small to large | Mixture-of-experts options |
Relative rankings shift with every release, so any claim that one family is simply better should be checked against current benchmarks and, more importantly, against the specific task.
A practical way to choose is to treat license terms, hardware budget and language coverage as the first filters, and benchmark scores only as the last one. Gemma tends to be attractive when a team already uses Google tooling, needs image input in a small model, or wants many languages from one checkpoint. A permissively licensed alternative may be preferable when legal review of custom terms would delay a project, and a hosted frontier model remains the better choice for complex multi-step reasoning that small models still handle poorly.
Deploying Gemma and its limitations
Gemma is deployed wherever data control, cost or latency matters more than peak capability. Common runtimes include Hugging Face Transformers, llama.cpp, Ollama, vLLM, Keras and Google Vertex AI, and Google publishes quantization-aware trained checkpoints that keep quality high at 4-bit precision.
- Hardware fit: the 1B and 4B models run on laptops and phones, while the 27B model generally needs a single high-memory GPU, especially at long context.
- Fine-tuning: parameter-efficient methods such as LoRA adapt Gemma to a domain on modest hardware.
- Limitations: smaller models hallucinate more, show weaker multi-step reasoning and carry less world knowledge than frontier systems.
- License obligations: redistribution must pass the use restrictions on to downstream users, which some legal teams treat differently from an OSI-approved license.
- Safety responsibility: once weights are self-hosted, the operator, not Google, is responsible for filtering and monitoring.
In simple terms
Gemma is a set of AI language models from Google that anyone can download and run on their own computer. Think of Gemini as a large restaurant kitchen that serves you through a window, while Gemma is a compact kitchen kit you can install at home.
The home version cannot cook every elaborate dish, but you control the ingredients, nothing leaves your house and there is no bill per meal. That is the main appeal of Gemma: privacy, low cost and freedom to adapt it.
A practical example
A regional law firm wants an assistant that summarizes contracts, yet its confidentiality rules forbid sending client documents to external services. The IT team installs the Gemma 3 12B instruction-tuned model, quantized to 4 bits, on one office server with a single 24 GB GPU, and connects it to the document management system through a retrieval pipeline.
Over two weeks the team fine-tunes a LoRA adapter on several hundred internal summaries written by associates. Lawyers now receive a draft summary of a 40-page agreement in under a minute, and every draft is reviewed by a person before use. The documents never leave the building, and running costs are limited to electricity and the server already owned.