# Open-Weight AI **Open-weight AI** refers to AI models whose trained parameters (weights) are publicly released, allowing anyone to download, run, modify, and deploy them without the original developer's permission or oversight. The most prominent examples are Meta's Llama series, Mistral's models, and — from Chinese labs — Alibaba's Qwen and DeepSeek's R1. The term "open-weight" is deliberately more precise than "open source." True open source implies access to training data, training code, and full reproducibility. Open-weight models release only the final product (the weights) — not the data they were trained on, not the RLHF annotations, not the training infrastructure code. You can use the model, but you cannot reconstruct how it was built. Meta's Llama, the most prominent "open-source AI" in public discourse, is open-weight with a restrictive commercial license — it is neither fully open source nor freely usable by competitors above 700M monthly users. ## Why Open-Weight Matters The open-weight vs. closed-model debate is the **access-vs-control fault line** of AI governance. The argument maps directly onto the KB's broader power-concentration analysis: **The case for open weights** (Meta's position, broadly shared by the open-source AI community): Closed models concentrate power in a few corporations. Open weights democratize access, enable independent safety research, allow governments and smaller organizations to build sovereign AI capability, prevent vendor lock-in, and create competitive pressure that keeps closed-model providers honest. Zuckerberg has framed Llama as a public good and a counter to the OpenAI/Google duopoly. **The case for closed models** (Anthropic and OpenAI's position, broadly): Once weights are released, the developer loses all control over how the model is used. [[Fine-Tuning]] can remove safety guardrails in hours. Open-weight models have been adapted for military targeting ([[Llama Military Adoption]]), disinformation generation, CSAM production, and other applications the original developer would prohibit. The argument: you cannot recall released weights, so open-weight release is an irreversible decision with potentially catastrophic downside. **The strategic play** (the cynical reading): Meta's open-weight strategy is also a competitive strategy. Meta cannot compete with OpenAI and Google on API revenue, so it commoditizes the model layer — making the foundational technology free while capturing value through the ecosystem (developers building on Llama become part of Meta's platform gravity). "Open source" is also a brand play: it positions Meta as the democratic alternative in an industry dominated by closed, for-profit labs. ## The Dual-Use Problem, Amplified Open-weight models make the [[Dual-Use AI]] ungovernable. When a model is closed (GPT, Claude), the developer can implement usage policies, monitor API calls, and refuse to serve military or surveillance customers. When a model is open-weight, the developer has no such leverage — the weights are downloaded and the developer's involvement ends. This is not theoretical. Llama weights have been downloaded by Chinese military-affiliated researchers, adapted for military applications by defense contractors, and fine-tuned to remove safety restrictions by hobbyists and criminals alike. Meta's acceptable-use policy prohibits these applications, but the policy is unenforceable once the weights leave Meta's servers. The U.S. government's position has oscillated: the Biden-era AI Executive Order focused regulatory attention on frontier models regardless of openness; the Trump administration has been more sympathetic to open-weight release as a counter to Chinese AI development, on the theory that widespread access to U.S.-origin models extends American AI influence globally — even if some of that influence flows to adversaries. ## KB Significance 1. **The ungovernable vector.** Open-weight models represent the point at which AI governance breaks down. You can regulate a closed API; you cannot regulate downloaded weights. Every governance framework documented in the KB — from the EU AI Act to U.S. executive orders — struggles with this asymmetry. Open-weight release is the structural reason that AI governance is harder than nuclear or biological weapons governance: the "materials" can be copied infinitely at zero marginal cost. 2. **The Meta strategy.** Meta's open-weight play is the most consequential competitive strategy in AI. By commoditizing the model layer, Zuckerberg is attempting to prevent OpenAI and Google from establishing the kind of platform monopoly that Microsoft achieved with Windows or Google achieved with Search. The KB's tech-oligarchy analysis includes Meta's open-weight strategy as a case study in oligarchs competing with each other for structural advantage. 3. **The geopolitical wildcard.** U.S. GPU export controls ([[GPU Export Controls]]) attempt to prevent China from training frontier models — but open-weight release gives Chinese labs access to U.S.-trained models that they can then fine-tune, distill, and build upon. The policy contradiction is stark: the U.S. restricts the hardware while American companies release the software. ## See Also - [[Foundation Model]] — The broader category - [[Dual-Use AI]] — The problem open weights amplify - [[Llama Military Adoption]] — The case study - [[Fine-Tuning]] — How open weights get repurposed - [[GPU Export Controls]] — The policy contradiction - [[AI Safety]] — The framework open weights challenge