A Week of Open-Weight Releases: The Good, the Bad, the Artificial
Three releases, three very different sets of promises kept and broken. A field report.
It was a strange week. Three labs shipped open-weight models within four days of each other, and the coverage treated all three as equivalent events. They weren't.
One release shipped with a genuinely permissive license and full training data provenance. Another shipped weights with a license that quietly forbade competitive commercial use — open in name, closed in function. The third was, generously, a fine-tune wearing a new model's clothes.
The good
Full provenance releases are rare enough that they deserve real credit when they happen. It lets the community actually audit training data concerns instead of taking safety claims on faith.
The bad, and the artificial
"Open-weight" has become a marketing term untethered from what it used to mean. We'd rather a lab be honest about a restrictive license than call it open and hope nobody reads the fine print.