What is Gemma?
Gemma is Google DeepMind's open-weight language model family. It is built on the same research and technology that powers Gemini, but released with open weights that anyone can download, fine-tune, and deploy. The latest generation, Gemma 3 (2025), is multimodal, supports a 128K-token context window, and is designed to run on a single GPU or TPU.
Developers get access to a powerful, well-engineered language model without being locked into a proprietary API. You can run it locally, customize it, or deploy it on your own infrastructure -- all under an Apache 2.0 license that allows commercial use.
The model family spans multiple sizes: from sub-2B on-device variants that run on phones, up to 27B-parameter models for workstation or cloud GPU deployment.
Key Features
Open Weights, Apache 2.0
Download the model weights, run them anywhere, fine-tune freely, and ship commercially. No API fees, no vendor lock-in, and a permissive license that covers both research and production use.
Runs on a Single GPU or TPU
Gemma 3 is sized to fit on one consumer-grade GPU, with smaller variants that even run on phones and laptops. The 4B model runs comfortably on an 8 GB laptop. You don't need a multi-node cluster to get serious work done.
Multimodal with 128K Context
Gemma 3 accepts text and images and handles a 128K-token context window, so it can reason over long documents, codebases, and rich multimodal inputs without chunking.
Available Across Developer Ecosystems
Pull Gemma from Hugging Face or Kaggle, deploy through Vertex AI on Google Cloud, or run locally with Ollama and llama.cpp. The same weights work across stacks -- no migration friction.
Sister Model to Gemini
Gemma shares research, tokenizers, and architecture patterns with the closed Gemini family. You get a lot of Gemini's engineering quality without giving up control of your data or weights.
How to Run Gemma
-
1
Pick a size that fits your hardware
Gemma 3 ships in multiple sizes: 1B (phones/edge), 4B (laptops, 8 GB VRAM), 12B (RTX 3090/4090), and 27B (A100/H100 or 2× consumer GPUs). Pick the largest size that fits in your GPU memory -- quality scales with parameter count.
-
2
Download the weights
Pull Gemma from Hugging Face, Kaggle, or Vertex AI's Model Garden. For quick local experiments, the easiest path is
ollama run gemma3:4b-- Ollama handles download, quantization, and the inference loop in one command. -
3
(Optional) Fine-tune on your data
For domain-specific quality, fine-tune Gemma with LoRA or full-parameter SFT using your own dataset. Hugging Face's TRL library, Google's Gemma cookbook, and Unsloth all provide ready-made fine-tuning recipes that run on a single GPU.
-
4
Deploy in your stack
Serve Gemma through vLLM or TGI for high-throughput inference, or via Ollama for local apps. The same weights work on Vertex AI, Hugging Face Inference Endpoints, Modal, RunPod, and other GPU clouds -- no vendor lock-in.
Use Cases
- --Local AI applications: Build AI features on your laptop or local server without sending user data to a third-party API -- great for offline tools, internal IDE assistants, and prototypes where you don't want to pay per token.
- --On-prem enterprise deployments: Deploy Gemma on your own infrastructure when compliance, data residency, or cost predictability rule out hosted APIs. Healthcare, finance, and government workloads that can't ship data outside the perimeter rely on open-weight models like Gemma.
- --Research and fine-tuning: Fine-tune for specialized domains (legal, medical, financial), low-resource languages, or task-specific behaviors. Open weights make it ideal for academic research and reproducible experiments where you need to inspect or modify the model.
- --Cost-controlled production inference: For high-volume tasks (classification, summarization, RAG), self-hosted Gemma can be dramatically cheaper than closed APIs at the same quality bar.
Pros and Cons
Pros
- +Open weights under Apache 2.0 -- free for commercial use
- +Runs on a single GPU/TPU, with on-device sizes for phones
- +Multimodal and 128K context out of the box (Gemma 3)
- +Available on Hugging Face, Kaggle, Vertex AI, and Ollama
- +Backed by Google DeepMind research -- same lineage as Gemini
Cons
- -Requires technical setup to self-host and serve
- -Larger variants need substantial GPU memory (24 GB+)
- -Generally trails the closed flagship Gemini on hard tasks
- -No official hosted chat product -- you build the UX yourself
Pricing and Model Sizes
Gemma weights are free under Apache 2.0. Your only cost is the compute you run it on. Below are the Gemma 3 sizes and their hardware footprints.
| Size | Hardware (4-bit quant) | Typical use |
|---|---|---|
| Gemma 3 1B | Phone / Raspberry Pi 5 | On-device chat, edge inference |
| Gemma 3 4B | Laptop (8 GB VRAM) | Local dev, lightweight assistants |
| Gemma 3 12B | RTX 3090/4090 (24 GB) | High-quality local chat, fine-tuning |
| Gemma 3 27B | A100/H100 or 2× consumer GPUs | Production inference, research |
Latest weights at ai.google.dev/gemma.
Alternatives to Gemma
- --Google Gemini: The closed, hosted flagship sister to Gemma. Pick Gemini when you want the strongest Google model with no infrastructure work.
- --ChatGPT: OpenAI's flagship hosted assistant. A natural baseline to benchmark Gemma against on general reasoning and writing.
- --Claude: Anthropic's hosted assistant with a strong reputation for long-context reasoning and safer tool use.
- --OpenAI Codex: Coding-tuned agent from OpenAI. Useful comparison if you'd otherwise fine-tune Gemma for code workflows.
- --Kimi: A long-context model worth comparing if your main reason for using Gemma is large-document reasoning.
Pick Gemma specifically when you need open weights you can self-host, audit, and fine-tune.
Who Should Use Gemma?
- 01.Developers building local AI apps: If you want a capable LLM running entirely on your machine -- no API key, no data leaving your laptop -- Gemma is the best-supported option from a major lab.
- 02.Teams with data sovereignty requirements: Healthcare, finance, and government teams that can't send data to hosted APIs get a production-grade model they can run inside their own perimeter.
- 03.ML researchers: Open weights make Gemma ideal for academic work, ablations, and experiments where you need to inspect or modify the model internals.
- 04.Cost-sensitive production workloads: Teams running high volumes of simpler tasks (classification, extraction, RAG) can save significantly by self-hosting Gemma vs. paying per-token API rates.
Not ideal for: Non-technical users, teams that need frontier reasoning without infrastructure work, or anyone who doesn't want to manage their own AI deployment.