GOVERNANCE STRUCTURAL SPLIT — AUGUST 2026
● Closed API (OpenAI/Anthropic/Google): Access controlled at lab level — 30-day window technically feasible
● Open-weight (Meta Llama, Kimi K3, DeepSeek V4): Weights downloadable after publication — cannot be restricted post-release
● Meta position: Held out of EO 14409 framework — White House pressed but no agreement reached
● Practical result: Framework covers closed-API frontier models only — open-weight models at same capability level are outside it
● Kimi K3 / DeepSeek V4 Flash: Both are open-weight frontier models — also structurally outside the framework
● CISA liability window: Information-sharing protections under CISA expire September 30, 2026 unless Congress acts
The EO 14409 pre-release framework creates a governance gap that is not about willingness to comply — it is about architectural feasibility. When Meta publishes Llama weights, anyone in the world can download them. A 30-day delay in publishing to give federal agencies a pre-release review window would mean a 30-day delay in publication — period. Meta cannot publish and restrict simultaneously. The same is true of Kimi K3's Modified MIT weights, DeepSeek V4 Flash's MIT weights, and Poolside Laguna S 2.1's OpenMDW weights. The framework as structured applies exclusively to labs that control access through an API or access program after the model is trained.
The incumbency concern: OpenAI, Anthropic, Google, Microsoft, and xAI are writing the capability threshold that determines which models require pre-release review. They are closed-API labs writing rules that structurally apply to them and their open-weight competitors differently. A threshold calibrated at a capability level that open-weight models will reach in one generation creates a compliance barrier that only closed-API labs can easily meet.
Last updated August 2, 2026. Related: EO 14409 full analysis → · Kimi K3 open weights →