Re: Open vs proprietary LLMs

On Tue, Aug 25, 2026 at 08:30:44AM -0500, Lucas Nussbaum wrote:
> # 2. Open Weight models.

You can break down Open Weight models into a number of different
subcategories.  Depending on the license of the weights, it might be
free for use for non-profit or companies with a revenue less than
$XX,XXXX or with more than XXX monthly active users.  So for example,
the Llama model is free so long have less than 700 million active
users.  If you have more, you need to get a special license from Meta.
The Llama license also prohibits the uses related to violence,
military/weapons, or other illegal activities.  Some licenses also
prohibit users from making changes to the model (so this might make
fine-tuning/post-training impermissible, although in practice this
might be hard to enforce).

Other Open Weight models might not have any such restrictions.  For
example the Gemma 4 model is released under the Apache 2 license,
which allows commercial users and doesn't have any discrimination
based in any person, group, or field of endeavor.

> Regarding ecosystems:
> There's a lot of vendor lock-in in the world of closed models: their
> providers try to control the full stack, from the model to the
> user-facing applications (coding agents, integration into office suites
> etc.)

In practice, most closed models are made available via endpoints that
are compatible with either the Olamma or Open AI network protocols.
This is what allows you to use a harness such as OpenCode under either
a local LLM or cloud-hosted models.  Some users / companies will also
have routers that will route simple queries to a local LLM server, and
fall back to cloud-hosted models (which might either be open models
running on a Cloud VM (e.g., Amazon Web Services), a ML-as-a-service
model to allow you to run a large Chinese open-weight model on a
cloud-hosted server in a data center (e.g., Amazon Sagemaker AI), or a
proprietary closed model (e.g., Anthropic Claude Opus).

The vendor lock-in comes in if you want to use a proprietary harness
(e.g., Claude Code versus Open Code) or a web service (e.g.,
https://gemini.google.com).  Some proprietary harnesses will allow you
to use an alternative LLM model, so there are varying levels of lock
in.

Some people have claimed that it's worthwhile to pay $20/month to use
Claude Code, even if you hook it up to local LLM if you have privacy
concerns (e.g., if you don't want to send your proprietary codebase to
Anthropic).  I'm not entirely sure it's a worthwhile tradeoff myself,
since Open Code is actually pretty good, but some people will claim
that.

> The GR discussion focused on the use of this technology for Debian
> contributions. When discussing usage, the level of freeness sounds like
> a more secondary concern than if the discussion was about inclusion of
> LLMs in the Debian archive. And it looks like, for ballot options
> proposers, the level of freeness was not that important compared to
> other factors/axes.

Given how quickly new models are available, and how large the more
sophisticated models are, I haven't seen much interest in including
LLM's in the Debian archive.  In practice most open weight users will
download LLM's on-demand from services such as HuggingFace.  This also
allows users to adjust which models to download based on how much
unified memory or VRAM their system might have.

Cheers,

						- Ted