By Roman Aglaganov
For a few years, Meta was the big tech company that gave its AI away. Researchers, start-ups and hobbyists built on its Llama models because they could download the weights, run them anywhere and change them. Its new Muse Wonder models, released on 12 October, come only through a paid API and $20 to $200 monthly subscriptions. As ITdaily noted, Meta said nothing at all about open weights.
The case for keeping it closed
I do not think every frontier model must be open. Once weights are published, safety measures can be stripped out, and nobody can recall them. Meta says Wonder has its strongest safeguards yet against cyber misuse, which only work if Meta controls the model. And the risks are not abstract: Anthropic’s red team showed this week that open models are getting closer to copying themselves across poorly secured devices.
There is also a business case. Meta spent heavily to build Wonder, and CNBC frames the launch as a test of whether users will pay. That is a legitimate thing for a company to want.
Why silence is the problem
What I object to is the silence. Thousands of projects were built on Meta’s promise that open models were central to its strategy. If that has changed, the people who relied on it deserve to know: will older or smaller models still be released? Under what conditions would Meta open a model again? What safety tests would a model need to pass?
What a better answer looks like
A clear policy: publish when, and on what evidence, Meta will release weights, as some labs already do for their own risk thresholds.
Keep the ladder: keep releasing capable smaller models openly, so researchers outside big companies can still study and build.
Open evaluations: if the model stays closed, let independent testers check claims like the MemoryBench lead, which are Meta’s own.
Open weights helped make AI research less concentrated in a few companies. Meta may have good reasons to step back from that at the frontier. It should say so, and say what it will still share.
