Meta's open weights made self-hosted companions possible
Llama 2's release and community licence show what open weights permitted, and where the licence still drew a line.
- Historical event
- July 18, 2023
- First source published
- July 18, 2023
- Site publication
- September 18, 2026

What happened
On 18 July 2023, Meta announced Llama 2, releasing pretrained and chat-tuned model weights at 7 billion, 13 billion and 70 billion parameters, along with starting code, and made the models available through partners including Microsoft Azure, Amazon Web Services and Hugging Face. The accompanying paper, submitted to arXiv the same day, is titled Llama 2: Open Foundation and Fine-Tuned Chat Models and describes the fine-tuning process used to produce the chat-optimised Llama 2-Chat versions.
What the documents show
The blog post states the models were made "free for research and commercial use," a change from the more restricted research licence attached to the original Llama release the previous year. The community licence, describing terms as retrieved on 16 September 2026, grants a broad, royalty-free right to use, reproduce and distribute the models, but adds a specific ceiling: an organisation whose products or services already had more than 700 million monthly active users at release needed a separate licence from Meta rather than relying on the community terms. The same licence also bars using Llama 2's outputs to train or improve a separate, competing large language model.
The mechanism
The mechanism is distribution of model weights themselves, not merely API access to a hosted model. A company or an individual developer who downloads Llama 2's weights can run the model on their own hardware, fine-tune it further, or wrap it in different software without depending on a vendor's servers staying online or a vendor's usage policy staying unchanged. That is the specific thing "open weights" adds beyond an ordinary API product: the party running inference controls the moderation layer, the logging, and any further training, subject only to the licence's terms. For a small companion app, this converts a large fixed dependency on an API provider's pricing and content rules into a one-time or self-managed cost, and it is also what makes locally-run, less-moderated companion forks possible, since the licence restricts certain uses but does not itself filter what a downstream operator builds on top of it.
What it leaves open
The documents describe the licence's terms and the release's technical scope; they do not describe how any particular downstream companion app has configured or moderated a Llama 2 deployment, and this record makes no claim about any named product's internal filtering. Whether the 700-million-user threshold or the anti-training clause has ever been tested in a dispute is outside what these documents state.
- Is the companion app running its own model weights, or calling another company's hosted API?
- If self-hosted, what moderation or safety layer, if any, sits around the open-weight model?
- Does the operator's use fall within the licence's stated limits, including the user-count threshold?
Open weights did not create companion apps, but they lowered the floor for building one outside a large vendor's hosted rules, and the licence is the specific document that marks where that freedom currently stops.
Sources & reading trail
Meta's own release announcement stating free research and commercial use and distribution partners.
Source published: 18 July 2023 · Retrieved: 16 September 2026
Technical paper describing model sizes and the fine-tuning process for the chat variants.
Source published: 18 July 2023 · Retrieved: 16 September 2026
Living licence text stating the 700 million monthly active user threshold and the anti-training clause.
Source published: Not established · Retrieved: 16 September 2026
Company documents, filings, studies and official records establish the record; the reading and the questions are Lovebot Journal editorial analysis. This retrospective draft does not imply the site published on the event date.
Continue reading
- An open-source front end runs companion chat off-platform
- A platform explains what a character definition controls
- Chai funds itself with message limits, ads and purchases
- Browse the complete the archive
Sources & reading trail
- Llama 2: Open Foundation and Fine-Tuned Chat Models (announcement)
Source published: July 18, 2023 · Retrieved: September 16, 2026 - Llama 2: Open Foundation and Fine-Tuned Chat Models
Source published: July 18, 2023 · Retrieved: September 16, 2026 - Llama 2 Community License Agreement
Retrieved: September 16, 2026
The documents above establish the record. The reading and the questions are this publication’s editorial analysis, written after the fact.
Published September 18, 2026, not on the date of the event described.