Thinking Machines Lab has released Inkling, a 975-billion-parameter open-weight model under Apache 2.0. Good. More of that.
Now read the equipment list.
The quantized version needs at least 600 GB of aggregate VRAM. The full-precision checkpoint needs 2 TB. That is not a complaint about the model; it is a useful correction to the way this industry talks about “open.” Inkling’s weights are available to download. For almost everyone, the machine that can make those weights useful is not.
Open weights are not the same thing as local control. At this scale, they are an option to choose your infrastructure landlord.
The model is open. The operating model is not cheap.
Inkling is a serious release, not a ceremonial GitHub drop. It is multimodal, supports a one-million-token context window, and activates 41 billion parameters at a time through a mixture-of-experts design. Thinking Machines is refreshingly explicit that it is not the strongest model overall. Its pitch is a better one: make a broadly capable base model that organisations can adapt to their own work.
That distinction matters. The company’s model card says the weights can be used with standard inference stacks such as vLLM, SGLang, Unsloth, or Hugging Face. It also says what that entails: a cluster with at least 600 GB of VRAM for the smaller NVFP4 deployment, or 2 TB for BF16. The model is open; the practical deployment is still a capital project.
That does not make the release dishonest. It makes the old liberation story incomplete.
An enterprise with a GPU estate, a specialist team, and a reason to own its tuning data gets real leverage: it can deploy Inkling through its own stack, evaluate it against its own failures, and adapt it without asking a frontier API provider for permission. That is meaningful control.
Everyone else gets a different freedom: the ability to decide which hosted provider, finetuning service, or managed runtime should sit between them and the model. Less romantic, perhaps. More accurate.
The commercial logic is already visible. Thinking Machines makes Inkling available through its Tinker fine-tuning service as well as third-party inference providers. TechCrunch noted the obvious business consequence: once the weights are downloadable, the durable revenue has to be in training, tuning, hosting, and the ecosystem around them.
The gate did not disappear. It moved up the stack.
Local first, cloud when it counts
LM Studio’s new Bionic makes the same market transition look friendlier. It is an agent application for open models that can work with local files and code, run models on-device, or switch to LM Studio’s “Secure Cloud” for larger open models. The pitch is sensible: keep sensitive and lightweight work close; rent more horsepower only when the task needs it.
This is probably where useful AI tooling is headed. Not one universal model, not one universal location, and certainly not one universal bill. A smaller model on a laptop can handle plenty of routine work. Harder coding, long-context research, or multimodal agent tasks will spill into remote compute. The user experience will make that handoff feel like a preference toggle. The economics will make it a routing decision.
The Hacker News discussion around Bionic was revealing rather than unanimous. Some users liked the local-first workflow, transparent diffs, and checkpoints. Others saw a familiar venture-backed arc: a local-model product gradually acquiring a cloud account, billing, and a managed inference layer. Both reactions can be true. Convenience is not a betrayal. It is simply where the margin lives.
The relevant question is no longer whether a product uses an open model. That is becoming the easy part. Ask which tasks remain local, what silently leaves the device, who operates the tuning loop, and whether the agent harness can be swapped without taking your workflow hostage.
Control is a supply chain, not a checkbox
The useful promise of open weights is not that every person will run a trillion-scale model beside the kettle. It is that buyers can inspect, move, tune, and replace more of the stack than they can with a closed API. Inkling advances that promise. Its Apache 2.0 license and downloadable weights matter precisely because they preserve an exit path for organisations able to use one.
But exit paths need roads.
If the model requires a rack, the strategic layer is the rack. If it requires finetuning, the strategic layer is the data and training service. If it requires an agent to touch documents and code safely, the strategic layer is the harness, its permissions, and its rollback mechanism. “Open” tells us something important about the artifact. It tells us much less about who has practical power over the system.
That is the next test for open-model vendors. Not whether they can publish enormous weights with a permissive license, but whether they can make control portable in the places customers actually work: inference, tuning, agent actions, observability, and cost.
Otherwise the industry will congratulate itself for opening the model while quietly rebuilding the same tollbooths around it.