Meta: the Llama family of open models
Try Meta’s Llama open models, ready to fine-tune and deploy on Crusoe infrastructure.
A dedicated endpoint on the GPU you choose
Llama models run on Self-Serve Deployments, so you get a dedicated endpoint instead of a shared queue. Pick from the latest NVIDIA and AMD hardware to match your latency and budget, with predictable GPU-hour billing.
Fast fine-tuning cycles
Smaller Llama models adapt quickly. Upload a dataset to Serverless Fine-Tuning, run a LoRA job with pre-configured defaults, and test a new version without standing up training infrastructure.
From post-training job to live endpoint in one place
Deploy a fine-tuned Llama checkpoint straight to a dedicated endpoint on the same platform where you trained it. No exporting weights, re-uploading, or stitching together separate training and serving providers.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
Meta launched its Llama family of open models in February 2023 and helped bring open weights into the mainstream, with one of the largest ecosystems of fine-tuning recipes, evaluation tooling, and inference engines built around it. Its AI work is now organized under Meta Superintelligence Labs, and the Llama Community License permits commercial use; a company whose products had more than 700M monthly active users on the model's release date must request a separate license from Meta. On Crusoe Cloud, Llama is available for Serverless Fine-Tuning and Self-Serve Deployments, running on Crusoe-operated infrastructure with no customer data sent to Meta. Request a Tailored Deployment for bespoke optimization and a dedicated endpoint.
Llama is open-weight rather than open source in the strict sense. Its weights are freely downloadable and can be modified and redistributed, but the Llama Community License adds conditions that OSI-approved open-source licenses don’t, including the 700M monthly active user threshold and an acceptable use policy. It also requires “Built with Llama” attribution, and models fine-tuned from Llama must have names beginning with “Llama.” Review the license and acceptable use policy before deploying.
Choose a smaller Llama model when you want fast responses, low serving cost, and quick fine-tuning cycles. Small Llama models run efficiently on a single GPU and make a strong base for focused tasks like classification, extraction, and domain-specific assistants. Larger Llama models deliver stronger reasoning and broader knowledge, but they need more GPU capacity and cost more to serve and tune.
Yes, both Llama 3.1 8B Instruct and Llama 3.3 70B Instruct are available for Serverless Fine-Tuning with a 131,072-token supported context. Fine-tuning is billed per 1M training tokens, where training tokens equal dataset tokens multiplied by epochs. Once the job completes you can deploy the checkpoint on a dedicated endpoint via Self-Serve Deployments or download the weights.
Llama is popular for fine-tuning because of its ecosystem. Nearly every major training framework, inference engine, and fine-tuning recipe supports Llama out of the box, and thousands of community fine-tunes show what the models can do. That maturity means fewer integration issues and more proven starting points, which makes Llama a reliable base for teams fine-tuning their first model.