Open Source & The Commons
The Weights Are Already Out There
Governments want to ban certain AI models, but unlike an app you can pull from a store, the core of an open-weight model — billions of numbers describing how it thinks — is already downloaded on thousands of hard drives around the world.
The Idea
When people talk about 'open' versus 'closed' AI, the word doing the most work is 'weights'. A model's weights are the numerical parameters baked in during training — the billions of values that determine how the system responds to any input. A closed model keeps these locked inside a company's servers; you interact with it only through an API, paying per query, never seeing the machinery. An open-weight model publishes those parameters for anyone to download and run on their own hardware. This is not the same as open source, though the terms often blur. True open-source software includes code, training data, and a permissive licence. Open-weight releases — like Meta's Llama 4 family or the models from DeepSeek and Moonshot — give you the trained parameters and usually the architecture, but often under licences with commercial restrictions, and rarely with full training data. The distinction matters, but for practical purposes the key shift is this: once weights are published, the model exists independently of its creator. The company cannot patch it, throttle it, or switch it off. For years, the obvious rejoinder was: so what? Closed models were simply better, so if capability mattered, you used a closed API and moved on. That calculation has collapsed. By 2026, the performance gap between open-weight and closed frontier models had narrowed dramatically — competitive across a wide band of everyday tasks, with open-weight models matching closed ones on some benchmarks at a fraction of the cost. The question is no longer which is smarter; it is who controls the infrastructure, who holds the data, and who pays the bill.
In the World
In July 2026, a Beijing-based startup called Moonshot AI released Kimi K3 — a 2.8-trillion-parameter open-weight model that it described as the largest ever announced. Within four days, the Trump administration was reportedly reviving efforts to restrict advanced Chinese AI models in the United States, citing cybersecurity concerns, with Kimi K3 as the immediate catalyst. The episode crystallised a tension that had been building for years. Chinese open-weight models — first DeepSeek, then Alibaba's Qwen family, now Kimi — had become genuinely competitive with US closed frontier models, and far cheaper to run. US graduate programmes were reportedly building on them; startups were self-hosting them because doing so kept data in-house and cost a fraction of the closed API price per token. But here is the enforcement problem that made the proposed ban almost farcical: open-weight models, by definition, already exist on hard drives worldwide. Analysts noted that an outright prohibition would be extraordinarily difficult to enforce in practice, because the weights can be downloaded, copied, and run locally with no ongoing connection to Moonshot or any other Chinese server. You cannot un-publish a file that is already in the wild. The geopolitical irony is sharp. The US leads in closed frontier capability — OpenAI, Anthropic, Google. China has increasingly owned the open-weight layer. And the tool Washington reached for to slow that down is a ban, deployed against technology whose defining feature is that it cannot be switched off.
Why It Matters
The open-weight versus closed debate is not really about software licensing. It is about where power accumulates in an AI-saturated world. Closed models create a form of infrastructure dependency: the company that owns the weights also controls access, pricing, safety filters, and the ability to shut a product down overnight. That is a rational trade for many teams — zero infrastructure to manage, instant access to frontier capability. But it means every product built on a closed API is, at some level, a tenant. Open-weight models shift that dependency onto whoever runs the hardware — which could be you, your government, or a foreign lab whose geopolitical interests diverge from yours. Europe has started treating open-weight models as strategic infrastructure, reasoning that a model you can self-host is one no foreign provider can switch off. The EU AI Act, which became fully applicable in August 2026, carves out partial exemptions for qualifying open-weight releases — but strips those exemptions away from any model deemed to pose systemic risk, regardless of licence. What the Kimi K3 episode made viscerally clear is that openness is not neutral. It is a power structure — one that distributes capability broadly and makes centralised control nearly impossible. Whether that feels like liberation or a threat depends entirely on who is trying to exercise the control.
A Question to Ponder
If the defining feature of an open-weight model is that no one can switch it off, does 'open' AI make the world more dangerous, more free, or is that simply the wrong question to be asking?
One reader's Thinkable for Tech Tuesday — a random draw from the library.
Start learning with Thinkable