Open Weights Do Not Automatically Give You Control

“Open weights” has become shorthand for independence. Download the model, run it yourself, stop paying rent to a vendor, stop worrying about deprecated endpoints. That is the pitch, and it is only half true.

Thinking Machines released Inkling, a frontier-scale model with downloadable weights. Teams can inspect it, fine-tune it, and in principle host it. Then you read the hardware line: the full checkpoint wants roughly 2 TB of GPU memory, and even the compressed version needs around 600 GB.

That is not a footnote. That is the whole story about what openness buys you.

Watch the Short Version

The short video version of this article is available here:

Watch the open weights portability video

What you actually receive

When a lab publishes open weights, you get the model files and a licence. You do not get the operating environment.

Think of being handed the blueprints for a factory. The blueprints are genuinely valuable - you can study them, modify them, take them to a different builder. You still need the building, the machines, the power contract, the technicians, and the safety systems before anything gets manufactured.

Weights are blueprints. Serving infrastructure is the factory. Most teams evaluating “open” models are pricing the blueprints and quietly assuming the factory comes free.

The portability gap

The gap looks like this in practice. A team picks an open-weights model specifically to avoid vendor lock-in. They ship on a hosted inference provider because that is the fast path. Eighteen months later the provider raises prices, changes terms, or sunsets the endpoint - and the team discovers they cannot actually move, because 600 GB of GPU memory is not something you improvise on a Thursday.

They had the right to leave. They never had the capability to leave. Those are different things, and only one of them shows up in a licence.

Worse, “we could self-host if we needed to” tends to function as a comfort story rather than a plan. Nobody has tested it. Nobody has priced it. Nobody knows whose job it would be.

Open does not mean local, cheap, or independent

Three claims get collapsed into one word:

• Open means you can obtain and modify the weights.
• Local means the hardware to run it exists inside your control.
• Independent means you can change providers without rebuilding your product.

Inkling is clearly open. For nearly every company, it is not local. Whether it makes you independent depends entirely on decisions you make after downloading it.

Four tests before you call a model your exit path

Run these separately. Passing one does not imply the others.

1. Rights. What does the licence actually permit - commercial use, fine-tuning, redistribution of derivatives, use in a product you sell? Read it rather than trusting the blog post.

2. Hardware. What does serving your real traffic cost at your real latency target? Price the compressed checkpoint, not the press-release configuration. Include redundancy, because one node is a demo, not a deployment.

3. Operating capability. Does someone on your team know how to run model serving, manage GPU memory, handle batching, monitor drift, and stay on call? If the honest answer is no, self-hosting is not an option you hold. It is an option you would have to hire for.

4. Exit path. How many days of engineering work stand between your current provider and a different one? If nobody has measured it, assume it is larger than you think. The useful version of this test is a rehearsal, not an estimate.

What this means for smaller teams

None of this is an argument against open weights. Open weights are strategically good: they cap a vendor’s leverage, they let you inspect behaviour, they give you a credible threat in a renewal conversation. A model you can obtain is better than one you cannot.

But the leverage is commercial, not operational. For a ten-person company, the realistic posture is: use hosted inference, choose models whose weights are obtainable, and keep a documented, costed, tested fallback - even if that fallback is a smaller model on cheaper hardware accepting reduced quality. Degraded service you control beats excellent service you cannot migrate.

That is the design we keep landing on when we build AI workflows at IndieStudio: abstract the model call behind your own interface early, so swapping providers is a configuration change rather than an archaeology project.

The operator takeaway

Open weights create options. They do not create control.

Control is an exit path your team can afford, staff, and has actually walked through once. Until someone has done that walk, “we’re not locked in” is a belief, not an architecture.

Source: Inkling: Our Open-Weights Model, Thinking Machines, 15 July 2026.