Re: Open vs proprietary LLMs

Lucas Nussbaum <[email protected]> Tue, 25 Aug 2026 10:07:41 +0200

Newsgroups	gmane.linux.debian.devel.general
Message-ID	<[email protected]>

On 25/08/26 at 07:56 +0000, Stephan Verbücheln wrote:
> On Tue, 2026-08-25 at 08:30 +0200, Lucas Nussbaum wrote:
> > It's useful to see LLMs as input data (or very large configuration
> > files) for inference engines such as vLLM or SGLang. They are not
> > executable on their own. They are actually similar to DNS blacklists:
> > you need to plug them into a piece of software that will make use of
> > them to provide a service.
> 
> Does that matter? The DFSG also applies to data included in the Debian
> repositories, not only to executable code. Many upstream programs
> include data that has to be removed before packaging because of license
> issues.
> 
> For generated data, it is preferred by free software projects that the
> code which generates the data is included in the source rather than
> data itself which is hard to manually study and modify.

There are two different sets of data in the context of this discussion:

1. the model itself (the weights). It's data generated from the training
data. It's clearly not the preferred form for making modifications (even
if you can post-train/fine-tune). Whether that could be included in the
Debian archive was the focus of the 2025 GR discussion[1].

2. the source code created with the assistance of an AI coding agent.
There, a human uses a tool (that indirectly leverages non-free data --
see (1)) to produce content in a form that is the preferred form for
further modifications to it. Whether such a process is acceptable within
Debian is the focus of the current GR.

[1] https://www.debian.org/vote/2025/vote_002

Lucas

signature.asc (application/pgp-signature, 833 B)