Google: the Gemma family of open models
Run Google's lightweight open models, built from the same research as Gemini.
Serve more users on the same GPUs
Crusoe’s inference engine is powered by MemoryAlloy™ technology, a cluster-wide KV cache that pairs with cache-aware routing to send each request to the node most likely to have its context ready. Gemma models handle more concurrent traffic at consistent low latency, without adding capacity to keep up.
Make Gemma your own
Run a LoRA fine-tuning job on your data with Serverless Fine-Tuning, deploy the result in one click, or download your weights in .safetensors format. With Gemma’s Apache 2.0 license, what you build is yours to run anywhere.
Go from first call to production with ease
Gemma is one of the few model families on Crusoe available across all three services: Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning. Prototype, tune, and scale to a dedicated deployment without switching models or platforms.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
Google builds its open-weight Gemma models through Google DeepMind, the Alphabet research unit formed in 2023 from the merger of DeepMind and Google Brain. Google backs them with first-party documentation, evaluation suites, and fine-tuning guides. In 2026, Google moved them to the Apache 2.0 license. On Crusoe Cloud, Gemma models are available for Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning, running on Crusoe-operated infrastructure with no customer data sent to Google. Request a Tailored Deployment for bespoke optimization and a dedicated endpoint.
Recent Gemma releases use the Apache 2.0 license, so commercial use, modification, and redistribution are permitted without fees. Earlier Gemma generations used Google’s custom Gemma Terms of Use, which included a prohibited-use policy. Check the license on each model card before deploying, especially if you’re working with an older Gemma release.
Gemma is Google’s family of open-weight models you can download, deploy, and fine-tune, while Gemini is Google’s proprietary model family available only through Google’s APIs. Gemma is built from the same research as Gemini but released at smaller sizes, so it can run on your own infrastructure, or with a provider like Crusoe, instead of only through Google.
Choose based on where the model will run and how much reasoning your task needs. Smaller Gemma models are built for efficiency, with fast responses and low memory requirements suited to high-volume or edge workloads. Larger Gemma models deliver stronger reasoning, coding, and instruction following, and typically support longer context windows for working across large documents. Start smaller and move up if quality falls short.
Many Gemma models accept both text and image inputs, so they can handle tasks like document understanding, chart analysis, and visual question answering alongside standard text generation. Gemma models are also trained on broad multilingual data, which makes them a strong fit for applications serving users across languages. Supported input types vary by model, so check each model’s details.
A fine-tuned Gemma model can match or outperform much larger general-purpose models on narrow, well-defined tasks, at a fraction of the serving cost. Gemma’s dense architecture and moderate size make fine-tuning faster and cheaper, and the tuned model runs with lower latency in production. This makes Gemma a strong choice for classification, extraction, domain-specific assistants, and other focused workloads.