Mira Murati did something unusual when Thinking Machines Lab released its first AI model this week. She admitted it was not the best model in the world.
“Inkling is not the strongest overall model available today, open or closed,” the company said in its announcement.
In an industry where every model release arrives wrapped in benchmark victories and carefully constructed claims of frontier performance, that sounds almost like an apology. It isn’t. It may be the clearest indication yet of what Murati is actually trying to build.
Inkling is the model Thinking Machines wants you to notice. Tinker is the business it wants you to build on.
That distinction matters because Inkling will inevitably be thrown into the AI model horse race. It will be compared with OpenAI’s GPT models, Anthropic’s Claude family, Google’s Gemini models and the rapidly expanding collection of open-weight models coming from China. There will be charts, rankings and arguments about whether Inkling is too big, too expensive, not powerful enough or better than expected.
Those comparisons have their place. But they risk missing the larger ambition behind Thinking Machines.
Murati does not necessarily need to build the one model that beats every other model. She is betting that enterprises will eventually care less about which model is best for everyone and more about which model can be made best for them.
Inkling Establishes the Foundation
Inkling is not a small undertaking. It is a 975-billion-parameter mixture-of-experts model, although only 41 billion parameters are active when processing a token. That architecture is intended to provide much of the breadth of a very large model without waking up its entire brain for every request.
According to Thinking Machines, Inkling was pretrained from scratch on 45 trillion tokens drawn from text, images, audio and video. It can reason across multiple forms of input, supports controllable reasoning effort and offers a context window of up to one million tokens.
The model was trained on Nvidia GB300 NVL72 systems, placing it firmly within the same capital-intensive AI factory model powering the rest of the frontier industry. Thinking Machines and Nvidia previously announced a multiyear partnership under which the startup expects to deploy at least one gigawatt of advanced Nvidia systems.
Inkling is a serious technical credential for a startup founded by OpenAI’s former chief technology officer. It demonstrates that Murati and her team can assemble the people, infrastructure, data and training systems necessary to create a major foundation model from scratch.
It also gives Thinking Machines its own model to optimize, improve and integrate deeply with the rest of its product portfolio. The company is already previewing Inkling-Small, a 276-billion-parameter model with 12 billion active parameters, suggesting Inkling is the beginning of a family rather than a single release.
Yet Thinking Machines is not presenting Inkling as the undisputed champion. It describes it as a broad and balanced foundation designed around multimodality, efficiency and adaptability.
That last characteristic is the one to watch.
Inkling does not need to be the best finished product. It needs to be an exceptionally good starting material.
The Chinese Model Comparison Misses the Larger Bet
One predictable interpretation is that Inkling represents an American response to the rising dominance of Chinese open-weight models.
There is some truth to that. Alibaba’s Qwen, Moonshot AI’s Kimi, DeepSeek, Z.ai’s GLM and other Chinese model families have become increasingly important alternatives to the closed systems offered by American frontier labs. Companies are using them to reduce costs, diversify their AI strategies and gain more control over how models are deployed and modified.
The United States has a strategic interest in producing competitive open-weight models of its own, and Inkling will be evaluated partly through that lens.
But treating Thinking Machines as merely another competitor to Qwen or Kimi badly undersells the company’s ambition. At some level, every model competes with every other model. At another level, those Chinese models can become raw material for the business Thinking Machines is building.
The evidence is already sitting in plain sight.
Bridgewater Associates did not wait for Inkling to begin working with Thinking Machines. Its AIA Labs used Tinker, Thinking Machines’ cloud-based post-training platform, to customize Alibaba’s Qwen3-235B model around Bridgewater’s financial-document workflows.
Bridgewater’s researchers were trying to automate six information-triage tasks drawn from the daily work of investors. The tasks included determining whether financial articles were relevant, identifying useful material in central-bank documents and separating original analysis from recurring boilerplate.
General-purpose models struggled to reproduce the judgment of Bridgewater’s investment professionals. Better prompting improved performance, but only to a point. The researchers concluded that expert judgment was too tacit and difficult to articulate fully in a conventional prompt.
Using proprietary examples labeled and corrected by Bridgewater experts, the team fine-tuned Qwen through Tinker. According to the joint report from Bridgewater and Thinking Machines, the resulting specialized model reached 84.7% average accuracy, compared with 78.2% for the best frontier model tested. The customized model reportedly made 29.8% fewer mistakes and reduced inference costs per task by 13.8 times.
Those results apply to Bridgewater’s particular tests. They do not prove that a customized Qwen model is universally superior to GPT or Claude.
That is precisely the point.
Bridgewater did not need a model that was universally superior. It needed a model that was better at exercising Bridgewater’s judgment on Bridgewater’s work.
And Thinking Machines did not need Bridgewater to use Inkling. Tinker created value with a Chinese open-weight model serving as its foundation.
In that example, Qwen was simultaneously an Inkling competitor and Tinker inventory. Thinking Machines could win even when its own model was not selected.
Tinker Is the Move Up the Stack
Tinker is a managed platform for training and customizing open-weight models. Developers write their training loops locally and maintain control over their data, algorithms and model choices. Thinking Machines handles the complex GPU work required to run the training.
The platform supports supervised fine-tuning, reinforcement learning, preference optimization and distillation. Its current catalog includes more than 28 supported models ranging from approximately 1 billion to more than 1 trillion parameters, including dense models and mixture-of-experts systems.
Thinking Machines describes Tinker as model-agnostic. That does not mean a customer can upload any open-weight model ever created and expect it to work automatically. The company curates the models it supports and says it will update that catalog as new models appear and older ones are superseded.
That curated approach could itself become valuable. The number of capable open models is growing too quickly for most enterprises to evaluate, deploy and optimize every option independently. A platform that makes models easier to compare, customize and replace could become more important as the underlying choices multiply.
Thinking Machines says customer data is used only to fine-tune that customer’s model and is not incorporated into the company’s own models. Customers can also download the customized checkpoints they create. Those provisions allow Tinker to present itself as an alternative to the walled-garden approach in which an enterprise sends its data into a centralized model it cannot inspect, modify or take elsewhere.
Inkling strengthens that platform without defining its limits. It gives Thinking Machines a native model designed from the beginning for customization through Tinker. It also attracts developers, generates attention and demonstrates what the platform can do.
Call Inkling a Trojan horse if you like, although there is nothing particularly hidden about the strategy. Inkling brings people through the gate. Tinker gives them a reason to stay.
The model is the showcase. The customization factory is the business.
This is also a classic move up the stack. As foundation models become more numerous and capable, much of the underlying intelligence will become interchangeable for ordinary enterprise workloads. One model will lead a coding benchmark this month. Another will offer better multimodal performance. A third will be cheaper or faster. Six months later, the rankings may look completely different.
Building a durable business solely on temporary benchmark leadership is going to become increasingly difficult.
Tinker moves Thinking Machines into the layer where enterprises create differentiation. The foundation model supplies general capability. The organization supplies proprietary data, workflows, experience, preferences and institutional judgment. The customization layer combines them into something competitors cannot obtain simply by subscribing to the same API.
That is where the value moves.
From Centralized Intelligence to Differentiated Intelligence
Thinking Machines gave this strategy an ideological foundation in its recent manifesto, “The Future Worth Building Is Human.” Drawing on Friedrich Hayek and Michael Polanyi, the company argues that productive knowledge is tacit, local, fleeting and dispersed among the people who acquire it through their work.
Central planning fails, the company argues, not because the planners lack intelligence but because the relevant knowledge cannot all be collected and understood in one place. Thinking Machines sees a similar limitation in the prevailing AI model: A small number of companies train centralized, general-purpose systems and expect them to serve the diverse needs of everyone else.
That argument echoes warnings we have heard from Microsoft CEO Satya Nadella and Palantir CEO Alex Karp. Enterprises risk surrendering part of what makes them valuable when they feed their accumulated knowledge into general-purpose models controlled by outside companies.
If every bank, software company, manufacturer and media organization uses the same models in the same way, AI may improve efficiency without producing much lasting differentiation. Worse, companies may help model providers learn from the expertise and data that previously distinguished their businesses.
Customization offers another path. Rather than moving all institutional knowledge into a universal intelligence, organizations can bring capable intelligence closer to their own knowledge.
There is still a contradiction in the Thinking Machines vision. Inkling was produced using an enormous concentration of Nvidia hardware, and Tinker performs its heavy training work on centrally managed GPU clusters. The company may be decentralizing the ability to shape intelligence, but the infrastructure required to produce that intelligence remains extraordinarily centralized.
Tinker does not eliminate the AI factory. It places an accessible customization layer in front of it.
That may be enough. Most enterprises do not want to build gigawatt-scale training clusters or become foundation-model laboratories. They want greater control over how models learn from their data and reflect the judgment of their people without having to operate the supercomputing infrastructure underneath.
Murati is betting that this middle ground will become a major market.
The model race will continue. Inkling will win some comparisons and lose others. Chinese open-weight models will keep improving. OpenAI, Anthropic and Google will push the frontier forward. New models will arrive claiming better reasoning, lower costs and longer context windows.
Thinking Machines does not need to win every round of that contest. If Tinker becomes the place where organizations turn those models into differentiated intelligence, the company can occupy a more durable and valuable position above the fight.
Inkling is the model Thinking Machines wants you to notice. Tinker is the business it wants you to build on.
And if Murati is right about where AI’s value is moving, winning the model race may matter far less than owning the place where every model becomes someone’s own.
