Vext Labs open-sources every retired Theron model. Retired foundation versions, retired LoRA adapters, and retired brain versions are released under Apache-2.0 roughly 30 days after they are superseded, in BF16 with no quantization. They are listed with a download link, the reason they were superseded, and the headline evaluation delta.
The Theron Agent SDK ships open under MIT: the core primitives, verifier kernels, and sample agents that let anyone build on the same patterns. The Stoa standard is open too, with the spec under CC-BY-4.0 and the runtime, SDK, and conformance suite under Apache-2.0.
What stays proprietary is the current model weights, the Hive fleet, and the council orchestration, all Vext-hosted. The principle is open at the edges, owned at the core: the story is public, the recipe is sealed.
Every retired Theron model under Apache-2.0 about 30 days after supersession, the Theron Agent SDK under MIT, and the Stoa standard. Current weights, the Hive, and orchestration stay proprietary.
The Agent SDK and the Stoa spec are open. The current model weights, Hive fleet, and council orchestration are proprietary and Vext-hosted.