Open Weights

What It Is

When people say an AI model has “open weights,” they mean the company that trained it has published the model’s parameters: the many billions of numbers, learned during training, that are the model itself. Download those numbers and you can run the model on your own hardware, study it, and change it, with no permission slip from the lab that made it. This is different from open-source software in the strict sense, and it is very different from the usual arrangement, where a lab keeps the weights private and rents you access through an interface you pay for by the call.

The distinction became concrete this month when Thinking Machines Lab released Inkling, its first model, as open weights under a permissive Apache 2.0 license. A lab that could have metered every request instead handed out the model. That choice is the whole difference between renting intelligence and owning a copy of it.

How It Actually Works

A trained model is, underneath, an enormous grid of numbers organized into layers. The weights are those numbers. Publishing them, usually on a hosting site like Hugging Face, lets anyone load the same large language model the lab built and run it locally. What open weights do NOT usually include is the training data or the exact recipe used to produce them, so “open weights” is a narrower promise than “fully open.” You get the finished engine, not the factory.

Owning the weights makes two things possible that a rented interface cannot. You can run the model offline, on your own machine, so nothing you feed it leaves your control. And you can fine-tune it: continue training on your own material until it answers in your voice and knows your domain. A rented model can be adjusted only as far as the vendor allows. An owned one is yours to reshape.

The Cost and Tradeoff

Owning is not free of burden. Open weights shift the running costs, and the responsibility, onto you. A rented service hides an enormous amount of engineering: the servers, the scaling, the safety filters, the uptime. Take the weights and you inherit all of it. The largest open models still demand serious hardware to run at full size, which is why most operators reach for a smaller variant or a community project that shrinks the model to fit a laptop. You also lose the vendor’s guardrails, which cuts both ways: more freedom, and more responsibility for misuse.

The honest tradeoff is control against convenience. Renting is easier and always current. Owning is more work up front and buys you something renting never can: a tool no one can price-hike, alter, or retire out from under you.

How TWO Uses It

The operator question is not “open or closed,” it is “which of my work should stand on ground I own.” Some tasks belong on a rented frontier model, where staying current matters more than control. Others, anything sensitive, anything you cannot afford to have change underneath you, anything core to how your business sounds, belong on weights you hold. TWO treats open weights the way a tradesman treats owning his own tools rather than borrowing the shop’s: a deliberate investment in independence, chosen where independence is worth the upkeep. This connects to model arbitrage, the discipline of matching each job to the cheapest model that clears the bar; owning a capable model outright is the far end of that same instinct.

Scott’s Take: Rent the frontier for what must stay current, but own at least one model outright, because the tool you cannot lose is the only one you can build a business on.

What to Watch Next

Watch whether the prominent labs keep some models open as a matter of strategy, or whether open weights settle into a second tier below the closed frontier models. Watch, too, how quickly community projects keep shrinking large open models to run on ordinary hardware, because that is what turns “open weights” from a headline into something an operator can actually put on their own desk.