IDA Edge, IDA AI, and IDA MoE research releases updated.
The public Hugging Face research trail now shows updated releases for three distinct bodies in the IDA family: a compact edge body, a larger interpretive body, and a mixture-of-experts body.
A family architecture makes the pressure visible.
The IDA program does not describe intelligence as one checkpoint with one responsibility. Its public research surface separates local inference, broader interpretation, and expert routing so each role can be examined on its own terms.
This separation also preserves a governance boundary. RegOS owns authority and execution. IDA can analyze, explain, challenge, escalate, and recommend, but an AI output does not silently become an attributable decision.
52.8M parameters
The compact paired body is documented for local inference, governed memory, recurrent state, local attention, and cognitive routing.
Open the IDA Edge page →1,011.8M parameters
The larger paired body is documented for text generation and governed interpretive work inside the IDA family architecture.
Open the IDA AI page →Mixture of experts
The mixture-of-experts release explores personality council behavior, escalation reserve, governed memory, and cognitive routing.
Open the IDA MoE page →The repository remains the release surface.
The company site adds public context and keeps the family relationship legible. Technical cards, artifacts, licensing, and gated access remain on the corresponding Hugging Face repositories.
The associated IDA family data repository documents a training substrate for evaluation, benchmarking, convergence, memory inheritance, governance, boundary-holding, and dissent preservation. This article does not infer evaluation results that are not present in the public release data.