Open weight model

In plain English

Open weight releases, from Meta's Llama to the DeepSeek, Qwen and GLM families, let researchers, companies and hobbyists run strong models on their own hardware. Supporters credit them with faster research, lower costs, transparency and independence from a few large companies. Many of the strongest open weight models of 2025 and 2026 come from Chinese developers.

For safety, release is irreversible. Once weights are public, safeguards can be removed by further training, and no lab can recall the model, monitor its use or shut it down. Testing before release therefore carries more weight. In 2026 US CAISSI called Z.ai's GLM-5.3 the most cyber capable open weight model released to date, about four months behind the US frontier. The Containment Board grades the developers of several leading open weight models, most of which score low on outside testing and shutdown commitments.

Why it matters for rogue SI

An open weight model with dangerous capabilities cannot be contained after release. If a model capable of going rogue is ever published, every kill switch is gone on day one.