Instance

Judge Instance on what a buyer can actually download, and the answer is nothing. The site states an intention and stops. For a benchmarking project, the deliverable is the eval harness and the metric definitions — without those, there is no way to tell whether temporal coherence is measured with optical flow, human raters, or a learned critic, and those choices change the numbers entirely. When would we pick this? Later, potentially. If Instance publishes a reproducible harness with a leaderboard that labs actually submit to, it becomes useful precisely because a shared benchmark lets you compare a new model against prior ones without rebuilding an eval stack. A YC-backed team with a stated focus on world models is a reasonable bet to eventually ship something. When do we pass? Right now, and for any deadline-driven work. If you have to report video-quality numbers this quarter, a static page cannot help you. The same applies if your team lacks the engineering time to integrate a research framework once it lands — nothing on the site suggests a managed or hosted version. In practice, the closest thing to an alternative is the existing published benchmark literature for video generation, which you can run today, warts and all. Those benchmarks have known criticism — prompt sensitivity, weak correlation with human judgment — but they exist, and a benchmark you can run beats a better one that has not been written. One caveat worth flagging: the news cycle around "instance" right now is dominated by unrelated stories about other projects and even cease-and-desist letters over alternative front-end instances. Expect search noise when you go looking for updates, and go directly to tryinstance.app instead. Our read is simple. This is a pre-launch research announcement,