Stay free if you only need free download of all gemma 4 model variants and commercial use permitted under the gemma license. Upgrade if you need managed deployment in vertex ai model garden with one-click endpoints and auto-scaling inference endpoints with per-second billing. Most solo builders can start free.
Why it matters: Self-hosting requires GPU infrastructure and MLOps expertise that smaller teams may lack
Available from: Vertex AI Hosted
Why it matters: Open-weights models from any lab, including Google, have historically scored below the largest closed frontier models on the hardest reasoning benchmarks
Available from: Vertex AI Hosted
Why it matters: Use is bound by the Gemma license terms, which include prohibited-use restrictions and are not OSI-approved open source
Available from: Vertex AI Hosted
Why it matters: Limited multimodal capabilities compared to Google's flagship Gemini models that handle native video, audio, and long-context vision
Available from: Vertex AI Hosted
Why it matters: Community ecosystem and third-party fine-tunes are smaller than Llama's, so off-the-shelf checkpoints for niche tasks may be scarcer
Available from: Vertex AI Hosted
Yes, Gemma 4 is released under the Gemma license, which permits commercial use, fine-tuning, and redistribution of derivative models. There is no per-token inference fee because you run the model on your own infrastructure or via a cloud provider's compute pricing. However, the license is not OSI-certified open source - it includes a prohibited-use policy covering things like generating CSAM, harassment, and certain regulated decisions. Most standard SaaS, enterprise, and research use cases are explicitly allowed.
Gemini is Google's closed, hosted frontier model family accessed through API and consumer apps; Gemma 4 is the open-weights sibling you can download and run yourself. Gemini Ultra-class models will generally outperform Gemma 4 on the hardest reasoning, long-context, and multimodal tasks because they are larger and use proprietary infrastructure. Gemma 4, however, gives you full deployment control, fixed compute costs, on-device options, and the ability to fine-tune freely. Many teams use both: Gemini for hardest queries and Gemma for high-volume, latency-sensitive, or data-sensitive paths.
Hardware requirements depend on the variant and quantization level. As a reference from prior Gemma generations: Gemma 3 1B ran on CPUs and phones, the 4B variant fit on a single consumer GPU (8 GB+ VRAM), the 12B needed roughly 16 GB VRAM, and the 27B required an A100 or equivalent (40–80 GB) at full precision or a 24 GB GPU with 4-bit quantization. Gemma 4 variants will have their own specific requirements listed on the model cards at release. Quantized GGUF builds via Ollama or llama.cpp typically cut memory needs by 2–4x. For production traffic, most teams deploy on Vertex AI, AWS, or Hugging Face Inference Endpoints rather than self-managing GPUs.
Gemma models are distributed through Kaggle, Hugging Face, Vertex AI Model Garden, and Google AI Studio, with Ollama and llama.cpp typically picking up community quantizations shortly after release. You will be asked to accept the Gemma license terms before downloading. The official source of truth is the Gemma page on deepmind.google, which links out to the supported distribution channels and provides reference code for inference and fine-tuning.
Google DeepMind has explicitly positioned Gemma 4 around advanced reasoning and agentic workflows, meaning it is trained and tuned to handle multi-step planning, tool calling, and structured outputs that agents depend on. For production agents, it is a strong open option, especially when you need predictable latency, on-prem deployment, or fine-tuning on private tool schemas. Compared to closed APIs like GPT-4 or Claude with mature function-calling, you may need to do more prompt and harness engineering yourself, but you avoid per-call costs and vendor lock-in.
Start with the free plan — upgrade when you need more.
Get Started Free →Still not sure? Read our full verdict →
Last verified March 2026