For the last three years, enterprise AI strategy has operated on a single assumption: the best model capability lives behind a closed API, and access to it comes with dependence on whichever vendor controls that endpoint. That assumption is no longer the whole picture. Thinking Machines Lab's release of Inkling, a 975 billion parameter foundation model under Apache 2.0, is the latest and largest entry in a pattern that has been building for two years: frontier-class capability is increasingly available as weights an enterprise can hold, inspect, and move, not just an API it can call. The headline is not that this makes AI cheaper. In some cases it makes ownership more expensive. The headline is that dependence on a single vendor is now a choice rather than a structural requirement.
Inkling is a mixture of experts transformer: 975 billion parameters total, 41 billion active per token, trained on 45 trillion tokens of text, image, and audio data, with a context window up to one million tokens. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, published the full weights on Hugging Face along with a quantized NVFP4 checkpoint for more efficient inference. Multiple outlets, including TechCrunch and InfoWorld, framed the release explicitly as a Western, openly licensed answer to the fact that the most competitive open weight models of the last two years, DeepSeek, Qwen, and Kimi, have come out of Chinese labs. For enterprises that were unwilling to run a Chinese-origin model in production regardless of license terms, that framing is not marketing. It is a genuine expansion of the option set.
The reason this matters goes beyond any single model. Open weights restore three capabilities that a closed API structurally cannot offer, no matter how good the model behind it is.
Inspection. You cannot see inside GPT or Claude. You can see inside Inkling. For regulated industries and any organization that needs to explain model behavior to an auditor, that is not a minor difference.
Modification. Fine-tuning a closed model happens on the vendor's terms, inside the vendor's infrastructure, and the resulting weights typically do not leave that infrastructure. Fine-tuning an open weight model produces an asset you own outright.
Portability. Inkling is already served by Tinker, TogetherAI, Fireworks, Modal, Databricks, and Baseten. The same weights, five different vendors. If one raises prices, changes terms, or degrades service, the workload moves. That is the capability a closed API cannot offer at any price: there is only one place to send the request.
None of these three requires an enterprise to own physical GPU hardware today. Every one of the hosting options listed above is itself running on GPU infrastructure, since inference at this scale has no CPU-only path. What changes is who operates that hardware and under what terms, not whether GPUs are involved at all. That distinction, who controls the compute versus whether compute is required, is the correct way to read the Postgres versus Oracle history. Most enterprises running Postgres today still consume it as a managed service from a cloud provider running on that provider's hardware. What changed when Postgres matured was not that everyone self-hosted. What changed was that Oracle could no longer price a customer as though no alternative existed. The alternative did not need to be adopted to be effective. It needed to be credible.
Here the analogy needs a correction, and it is an important one for anyone building a business case. Postgres runs on a laptop. The marginal cost of self-hosting an open source database has been close to zero for two decades. Inkling does not share that property, and no model at this scale does, open or closed. The full precision checkpoint requires roughly two terabytes of aggregated GPU memory. Even the quantized version requires an estimated 600 gigabytes. This is not infrastructure a team spins up on a whim. It is a genuine capital and operating decision, on the scale of enterprise data center infrastructure, not a side project, and it applies whether the enterprise runs that hardware itself or pays a hosting vendor to run it on their behalf.
This is why on-premises versus cloud is the more accurate analogy than open source versus SaaS, and it is worth being precise about the distinction. Open source software commoditized the cost of the software itself, leaving infrastructure as the only remaining cost. Open weight foundation models do not commoditize infrastructure at all. GPU cost is present in every deployment path, closed API or open weight, self-hosted or vendor-hosted. What open weights commoditize is the choice of who runs that infrastructure for you. An enterprise adopting Inkling today will, in nearly every case, consume it through one of the five hosting vendors listed above rather than running it on owned hardware. The freedom being purchased is not freedom from GPU cost. It is freedom from a single point of dependence, exercised through the ability to move the same workload to a different vendor's GPU infrastructure without renegotiating the model itself.
That distinction should shape how a technology leader evaluates the decision. The right question is not “can we avoid the GPU cost.” The GPU cost exists regardless of vendor, open or closed. The right question is “does the option to leave exist, and what would it cost to exercise it.” That question has a much better answer with an open weight model than with a closed one, even when the underlying hardware requirement and the day-to-day deployment look identical.
One detail in the Inkling release deserves a direct correction because it gets conflated in coverage of the launch. Reporting has noted that Inkling's training process partly used synthetic data derived from Moonshot AI's Kimi K2.5, a Chinese model, even as Thinking Machines Lab positions Inkling as a Western alternative to Chinese open weight models. That is a fact about training data provenance, a step that happened once, months before release, when the weights were built. It is not a fact about where your data goes when you use the model today.
Runtime data flow is determined entirely by which inference provider processes your requests, not by which datasets trained the model's weights. An enterprise self-hosting Inkling, or routing through a US-based provider such as TogetherAI or Databricks, has the same data residency posture it would have with any other model hosted by that provider. Training lineage and inference data flow are two separate questions, and enterprise risk assessments should treat them separately. Conflating them produces either false comfort or false alarm, depending on which direction the conflation runs.
The freedom argument is real, but it has limits worth naming plainly.
GPU cost is not optional at any point in this equation. Whether an enterprise self-hosts or routes through a vendor, someone is paying for GPU infrastructure at a scale most organizations have never had to budget for internally. Open weights do not change that fact. They change who bears it and under what terms.
Most enterprises will not self-host in the near term. The hardware requirement puts direct ownership out of reach for all but the largest organizations, which means most of the benefit today is portability across hosted providers, not independence from hosting entirely.
Fine-tuning still requires capability the organization may not have. Owning the weights is not the same as having the MLOps expertise to productively modify them. That expertise is a real, ongoing investment, not a one-time cost.
Quality parity is not guaranteed. Open weight models are closing the gap with closed frontier models, but closing a gap is not the same as eliminating one. Model selection should still be driven by evaluation against the actual task, not by license type alone.
License diligence still matters. Apache 2.0 is permissive, but enterprise legal and procurement review should confirm terms on each new release rather than assuming all open weight models carry identical obligations.
Open weight models did not make frontier AI free, and they did not remove GPUs from the equation. What they did is put a price on independence that an enterprise can now choose to pay, where before there was no option to buy it at any price. That is the actual shift, and it is the one worth building strategy around.
Let's talk about where you are today and where you want to go - our experts are ready to help you move forward.
