Beyond OpenRouter
This library uses official model documentation, publisher repositories and lifecycle notices. A model does not need an OpenRouter listing, usage history or a place on the map to be included. Map activity is never used to invent specifications.
Scope, and its limits
We cover named language, coding, reasoning and text-producing multimodal models offered through APIs or official downloadable weights, including historical versions. Embeddings, rerankers, reward models, media-only generators, adapters, intermediate checkpoints and third-party repackagings are outside this edition. Runtime and quantization downloads are folded into aliases where an exact parent is established. Base and instruction-tuned releases remain separate.
This is a broad publisher inventory, not a claim to contain every community fine-tune or every AI model ever created. “Source listed” confirms a publisher listing; it does not guarantee API access, an unrestricted license, or a working download in your region. Deprecated and retired models remain searchable through the status filter.
Coverage is not a quality score
“Documented specifications” means at least one token limit or cutoff is supported by an official source. It does not imply every field is complete, independently tested or editorially recommended. Gated or incomplete documentation leaves visible gaps. Each field links its evidence. A missing field is not zero and does not mean a capability is absent.
What the expanded pages add
Every model has a permanent page with answers supported by its recorded evidence. 1,386 records include a freshly checked official checkpoint identity, revision and available repository details. Stored parameter counts, architecture labels and pinned configuration values describe the distributed weights; they are not tested hardware requirements or hosted API limits.
1,679 catalog pages meet our current search-indexing evidence gate. The remaining 20 are accessible reference pages that we ask search engines not to index until their evidence is stronger. Previously approved editorial profiles remain available on their own reviewed merits. The gate counts sourced decision facts, not word count or copied boilerplate. A reviewed model card can qualify through exact-checkpoint implementation, training and use details even when a publisher does not disclose a parameter count. A retired-model reference can qualify through its dated lifecycle record, replacement guidance and platform consequences. A focused pricing reference can qualify through exact-model rates, currency, billing units and deployment conditions, with a worked cost example. These source reviews are bound to the facts they checked; changed facts require a new review.
Browse every permanent model page →
New releases and corrections
Our release-monitoring workflow checks official inventories and catalog pages for changes. A detected change is a research candidate, not proof of a launch. New entries require a verified model identity and official evidence; aliases, changed downloads and page formatting are checked before anything is added. The checked date above describes this published edition, not a promise of real-time coverage.