These resources provide unrestricted access to model parameters, allowing you to inspect, modify, and host systems on your own infrastructure. Whether you are building localized language processors or training custom vision pipelines, these collections offer full transparency into the underlying architecture. Prioritize your selection based on the specific hardware requirements of your server and the depth of control you need over the final output behavior.

Open 32B model: only 18 of its 72 layers keep a KV cache