Meta's Llama license shows what open weights really allow

Start here

Most assistants run on someone else's computer: you send a request, a server processes it, and a reply comes back. 'Open weights' describes a different arrangement -- the trained model file is published for anyone to download and run on their own hardware, with no round trip to a vendor's server. Meta's 24 February 2023 post, Introducing LLaMA: A foundational, 65-billion-parameter large language model, released the model in four sizes, from 7 billion to 65 billion parameters, trained on '1.4 trillion tokens' at the largest scale. For an ordinary user, that shift changes both what a conversation costs and where its data goes.

What the documents say

The original release was 'under a noncommercial license focused on research use cases,' with access granted case by case to academic researchers, government and civil society groups, and industry labs, not the public. The post also names limits current chat products still carry: the model shares 'bias, toxic comments, and hallucinations.' Five months later, Meta and Microsoft Introduce the Next Generation of Llama, dated 18 July 2023, changed the terms: 'Llama 2 is free for research and commercial use,' citing 'more than 100,000 requests for access' to the original as evidence of demand. The Llama 2 Community License, retrieved 16 September 2026, grants a broad license but with one checkable trigger: an organization above '700 million monthly active users' in the preceding month must request separate permission from Meta before using it at all.

Check this

Two checks follow. First, 'free' and 'open' are not the same as unrestricted: read the license's actual conditions, including that 700 million user threshold, before assuming a commercial product can rely on it without permission. Second, if privacy is the reason you want a locally run model, verify the claim: a model running fully offline should show no outbound network activity for your requests, checkable directly, unlike a cloud assistant where you must take the provider's data-handling statement on trust.

What holds and what fails

This is an editorial reading of what 'local use' buys you: the privacy and cost argument for open-weight models holds up because it is independently verifiable -- no server call means no company logging that request, and you can confirm this yourself rather than trust a policy page. It fails as a claim of total freedom: the license still restricts very large organizations, requires an attribution notice on redistribution, and forbids using outputs to improve a rival model. Openness describes who can run the model, not the absence of rules governing it.

• Before calling a local model 'private,' check it truly runs offline with no outbound requests for your data.
• Read the license's specific thresholds, like the 700-million-user trigger, rather than assuming 'free' means unconditional.
• Compare running a model on your own hardware against a subscription before assuming local is automatically cheaper.

The gap between a research-only license in February and a commercial one by July is worth keeping in mind: what a vendor calls 'open' can change in months, and the license text, not the announcement's tone, actually governs your use.