Inkling is Mira Murati's first release since leaving OpenAI, and it arrives open-weight on purpose. It is a mixture-of-experts system, 975 billion parameters with about 41 billion active on any given task, trained across text, image, audio and video. The company is unusually candid about what it is not. In its own briefing it calls Inkling not the strongest model available today, closed or open. That is not modesty for its own sake. It is a positioning move. Inkling is not sold as a finished chatbot to compete with the frontier. It is sold as a starting point, something an organization fine-tunes on its own knowledge through Tinker, the company's customization platform. The bet underneath the whole release is that a model a company can shape for itself will outwork the one everyone rents.
The clearest evidence for that bet did not come from a benchmark leaderboard. It came from a project with Bridgewater Associates. Researchers took an existing open-source model and trained it further on the fund's own financial expertise. The result scored 84.7 percent on financial reasoning tests, beat the top proprietary models, and cost roughly a fourteenth as much to run. The number is the two companies' own evaluation, not an independent one, so hold it loosely. But the shape of the result is the point, and the shape is what matters. A generic base plus a specific body of knowledge produced something that outperformed models many times more expensive on the task that was actually being measured.
This is a quiet inversion of where value has sat for three years. The differentiator was access to the best general model, and access was rented by the token. Now the base is becoming a substrate, close to a commodity, available open-weight from more than one lab. What sits on top of it is where the advantage moves: the proprietary knowledge composed into it, the corrections and context and hard-won practice that only your organization holds. Satya Nadella named the flip side of this the same week, warning that enterprises on proprietary models pay twice, once in subscription and again in the business knowledge their prompts hand over to be absorbed into the next version. Hugging Face's Clem Delangue put the same split more plainly: frontier models for experimentation and the highest-value work, private or open models for most production.
The base model is a commodity. The knowledge you draw into it is the asset.
None of this is free, and the honest version of the argument says so. Owning a model means owning its safety, its upkeep, and the machine-learning talent to tune it well. The renter buys simplicity and a lab's guarantee. The owner buys control and inherits responsibility. Fine-tuning is not a button. It is a capability a company staffs for, and the moment a model is customized, the safety of that customization stops being the lab's problem and becomes yours. For many organizations the rented general model will remain the right call, and that is a fine answer. It stops being a fine answer when nobody ever asked the question.
The question is a sorting exercise, and it is architectural before it is technical. Take the processes that now run on a general model and separate them. Which ones lean on knowledge that is specific to your organization, the kind a specialized model would learn and a rented one never will? Those are the candidates to own. Which ones are generic enough that a rented model is already good and always will be? Leave them rented. The gain from specialization is not evenly distributed across a business. It concentrates in exactly the places where your practice differs from everyone else's, and those are the places worth pouring knowledge into a model you control.
We wrote earlier in these pages that the model is rented, that treating a vendor's model as a swappable commodity behind a boundary you own is the sober way to build. This is the other half of that same thought. Renting was the right posture when the base model was the differentiator and owning one was out of reach. The base is no longer the differentiator, and owning a specialized one is now a real option rather than a research luxury. The model can be rented, still, for most of what a business does. And now, where your knowledge is the edge, the model can be owned. Deciding which is which, on purpose, is the work.
Thinking Machines released Inkling as an open-weight model built to be fine-tuned, with a Bridgewater fine-tune it says beat proprietary models at a fraction of the cost. Reported by TechCrunch.