Open vs. closed weights: the licence divide that shapes who controls frontier AI

What it is

The "open vs. closed" beat tracks whether frontier AI labs publish the trained numerical parameters of their models, known as weights, for anyone to download, run locally, and modify. An open-weight release removes the lab as a gatekeeper: any developer, company, or government can deploy the model on its own hardware, fine-tune it on proprietary data, and build products without a cloud subscription or an export licence from a foreign government. A closed-weight model is accessible only through an application programming interface controlled by the lab, which can restrict access by nationality, revoke it, or adjust pricing unilaterally.

The distinction carries direct world-news significance. Countries under US chip or software export controls cannot legally access US-gated closed models, but open-weight models, once published, are unrestricted. For enterprise buyers, open weights undercut closed-model API pricing by three to six times on comparable tasks. For governments, open weights raise a governance question that closed models do not: proliferation is irreversible once weights are downloaded at scale.

History

Meta's Llama 2 (July 2023, 7 billion to 70 billion parameters) was the first large open-weight release from a US frontier lab, seeding an ecosystem of fine-tuned derivatives. Mistral AI, founded in France in May 2023, released its first open-weight models under Apache 2.0 in September 2023. The inflection came in January 2025 when China's DeepSeek released R1: an open-weight model matching US frontier performance at an estimated US$5.6 million in training compute, triggering a US technology stock sell-off. By August 2025 the OECD had published a formal policy paper on AI openness trade-offs. In May 2026, the Linux Foundation released the OpenMDW-1.1 licensing standard, adopted immediately by Nvidia, providing a legally coherent rights-grant for model weights that Apache 2.0 could not. Days later, the G7 Digital and Technology Ministers replaced the binary open/closed classification with a spectrum framework.

Current state

As of mid-2026, Stanford HAI's 2026 AI Index puts the capability gap between the best closed and the best open-weight frontier model at 3.3 percent on major benchmarks, up from 0.3 percent in August 2024. The UK AI Security Institute estimated the open-weight lag at four to eight months. Two Chinese labs are driving the open side. DeepSeek released V4-Pro open-weight in April 2026 (1.6 trillion parameters, 49 billion active, one-million-token context). Zhipu AI released GLM-5.2 in June 2026 (744 billion parameters) under an MIT licence carrying no regional restrictions, which beat GPT-5.5 on long-horizon coding benchmarks at roughly one-sixth the per-token cost. On the closed side, US labs OpenAI, Anthropic, and xAI retain all weights as proprietary; Google DeepMind publishes some weights while keeping top-tier systems API-only.

Relationships

The open-vs-closed sub-beat has no formal roster subjects, but it threads through every story in the compute-frontier cluster. DeepSeek V4 preview narrows the gap with open-weight MoE models and Zhipu AI releases GLM-5.2 open-weight under MIT licence, beating GPT-5.5 on coding at one-sixth the cost are the two live open-weight events defining current dynamics. Open weights gain at the bottom as the frontier closes at the top tracks the split quarter by quarter, documenting the tier structure where open models are near-frontier but the very top remains closed. Open-weight licensing gets a standard, and the G7 a spectrum covers the governance layer: the OpenMDW-1.1 licence standard and the G7 spectrum framework that replaced the binary classification in June 2026. Geopolitically, the beat connects to US AI export controls: Chinese open-weight labs using MIT or Apache 2.0 reach markets where US closed models are gated. On AI safety, open-weight releases pose an irreversible proliferation risk no lab can patch once weights are distributed at scale.

What to watch

Whether the US government brings MIT-licensed open-weight Chinese models under export control frameworks, which would require defining "open weights" in regulation for the first time. Whether any lab open-weights a model that leads every major benchmark, closing the current 3.3 percent gap. How national regulators operationalise the G7 spectrum framework, in particular whether it changes government AI procurement standards. Whether OpenMDW-1.1 becomes the de facto open-weight licence and whether Alibaba (Qwen) and Meta adopt it. The pace of open-weight adoption in markets where US models face access restrictions, where GLM-5.2's MIT licence and DeepSeek V4's Hugging Face availability already offer a frontier-class alternative.