What counts as a release on this site

"Release" is a sloppy word in AI. A keynote demo, a waitlist, a research PDF, and a generally available API all get the same headline treatment. They are not the same event. This desk uses a narrower definition so the archive stays useful six months later.

It counts

A versioned model that someone can call or download. A new API model ID, a new open-weight checkpoint with a license, or a documented upgrade to an existing ID (context, modalities, tool use) is a release.

A documented serving or tooling change that changes how you run models. A vLLM release that adds a scheduler, an Ollama version that ships a new engine, or a Transformers release that changes default tokenization is in scope. The point is operational: a practitioner can do something today that they could not do yesterday.

A retirement or breaking change. Deprecations belong here. If an old model ID is going away, that is more useful than another "we are excited" launch.

It does not count

A teaser. "Coming soon," "in research preview with no docs," or a social clip without a model card is not a release. We can mention it later when the docs exist.

A rumor. Two anonymous posts and a screenshot of a playground do not clear the gate. If a lab later confirms it, we write from the confirmation.

A wrap-up of wraps. If the only new information is that five other blogs already wrote about the same blog, we skip.

A paper with no artifact. A research result can run under the research tag when there is a paper plus code, weights, or a reproducible eval. A PDF alone is usually a skip unless the lab frames it as a product drop.

How we label posts

• releases — new or upgraded models, APIs, and deprecations
• research — papers with artifacts
• tools — serving, inference, and developer-stack ships
• industry — desk standards, policy, and process (like this post)

The tag is a filing cabinet, not an SEO costume. A post gets one primary tag in the UI even if the work touches two rooms.

Why this is strict

Scaled, unoriginal pages are how both search and ads programs lose patience. The fix is not a disclaimer. The fix is fewer pages, each one tied to a source you can audit. If we only publish when something moved, the site stays small on purpose.

If you want us to watch a source we missed, say so in a comment on this post and name the URL.