Loading page…
An open-weight model family: the hosting organization, application controls and exact release shape the system users encounter.
Llama names a model family, not one consumer service with universal settings. A developer can build very different applications around related weights. Ask who runs the service, what changes were made, and who handles logs, incidents and user requests.
Meta’s Llama 3.1 card documents an instruction-tuned model release, its evaluations, intended use and community license. These details belong to that release; they are not a substitute for the documentation of a later model or third-party fine-tune.
Source: Meta · Llama 3.1 model cardThe Open Source Initiative’s definition also considers freedoms to use, study, modify and share, along with access to relevant components. Read the actual terms before equating downloadable weights with unrestricted use.
Source: Open Source Initiative · Open Source AI Definition · 1.0Our interpretation: self-hosting gives an operator control over its deployment, but also leaves that operator responsible for authentication, data handling, moderation and updates. Those responsibilities do not disappear because model weights are available.
Two document assistants use the same Llama checkpoint. One searches a single approved folder; the other indexes the entire organization. Their exposure differs even before you compare their answers. Evaluate the system around the model as carefully as the model itself.
This page does not certify any third-party Llama service, declare every Llama release equivalent, or provide a license interpretation for a particular business arrangement.
How to read an AI safety evaluationThe sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
A concrete open-weight release: intended use, deployment considerations, evaluations and license terms.
One explicit definition of openness, covering freedoms and access to more than model weights.
Practical guidance on tool permissions, memory isolation, oversight and agent failure handling.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence