Open models, rebuilt in public.

We take open-source AI models and attempt to bring them to a fully reproducible state: lineage captured at runtime, registered publicly, rebuild attempted, receipts published. Models are judged not by their paperwork, but by whether they actually rebuild.

The Reproducible AI List — one page per model, with what was attempted, what broke, what it cost, and the lineage hash to reproduce it yourself.

Every rebuild is performed by a clean room — a fresh instance of an agent, no memory of prior rebuilds, working from the model's published materials and nothing else — running every command under roar, which captures lineage as the work happens and registers it on GLaaS, so every row links to a DAG you can inspect. Most repos in this org are forks: rebuild targets that needed patches, each carrying a diff against upstream. The rebuilt set is a selected set, not a random sample of open-source ML — aggregate numbers describe these models, chosen this way, and blockers are categorized to describe ecosystem gaps, never to grade individual authors.

Built one of these models? If we got your row wrong — or a newer release fixes a blocker — email reproducible-ai@treqs.ai and we'll re-run it. Same address to suggest a model for the list.

Run by TReqs