Llama.

Model as the basis of your application

Choose Llama when control over an AI feature's hosting and behavior matters. Developers can obtain suitable weights, configure request processing and embed the model in their own system. A ready-made chat, document search and access management do not appear automatically: they must be built around the model.

Before choosing, define the task and constraints. Brief request classification and complex conversations about long documents may need different configurations. A larger model will not necessarily be more practical if it cannot be hosted on the available infrastructure.

Options vary in capabilities

The official repository lists generations, licenses and instructions. The family has text and multimodal options, including models with image processing. The set of input data depends on the specific release: the presence of a visual mode in one model is not transferred to all others.

Distinguish a base model from an instruction-tuned model. Base weights require further preparation for the expected user-facing behavior. For ordinary conversation, check the recommended message format and model type in the official model card.

Example: an assistant for internal instructions

A team can build an answering system using its own policy base. First retrieve relevant passages, then give the model only the selected excerpts and question. Require references to internal documents and allow an honest "insufficient information" response.

Test tasks with conflicting and missing information. If a rule is not found, the system must not present a guess as a company directive. This is an architecture and testing example, not a claim about a ready-made feature in any downloaded checkpoint or a test we performed.

Using your own hardware still involves costs

The official instructions describe running on a GPU and reduced-precision options. Requirements can vary significantly; a long context uses additional memory. Before deployment, assess not only loading the weights but also concurrent requests, storage, updates and monitoring.

Local hosting lets you build a controlled processing environment. It does not guarantee privacy by itself: check logs, the app's external connections, user permissions and document-upload paths. Secrets must not appear in publicly accessible diagnostic logs.

License and result verification

The Llama open weights are distributed according to their license terms and conditions of use. Read the documents of the selected generation, including distribution requirements and acceptable scenarios. The availability of downloads is not equal to the absence of any restrictions.

Check your own application

The model can make mistakes, reproduce data biases and confidently fill gaps in information. Test your application's quality on its actual tasks; measurements published by the developer do not replace that testing.

Suitable tasks

• Build your own text AI service
• Build answers from verifiable document passages
• Run a suitable model in your infrastructure
• Compare configurations for code and multimodal input