Thinking Machines Lab has introduced Inkling, its first internally developed artificial intelligence model, as the company pushes a vision of AI that businesses can adapt to their own needs.
Founded by former OpenAI chief technology officer Mira Murati, the startup is taking a different approach from companies that primarily offer access to closed, general-purpose models. Inkling is an open-weight model, which means developers and organizations can download its model weights, modify the system, and customize it for specific applications.
The release gives the industry its clearest look yet at the technology Thinking Machines has been developing. It also places the startup in the growing competition between proprietary AI platforms and models that organizations can operate and refine more independently.
Inkling uses a mixture-of-experts architecture with 975 billion total parameters. However, the model activates only about 41 billion parameters for an individual task.
This design allows a large model to call on specialized groups of parameters instead of using the entire system for every request. The approach can reduce computing requirements while maintaining broad capabilities across different types of work.
Thinking Machines says Inkling was trained on 45 trillion tokens covering text, images, audio, and video. Although the model was trained to reason across multiple forms of information, its current outputs are limited to text.
Those outputs can include ordinary written responses, computer code, structured data, and styled digital content. Future versions could potentially expand the range of content the system can produce.
One of Inkling’s notable features is its focus on calibrated responses.
AI models frequently produce confident answers even when they do not have enough reliable information. Thinking Machines says Inkling is designed to recognize uncertainty and signal when it may not know the correct answer.
This could be valuable in business environments where an incorrect answer presented with confidence can lead to financial, operational, or legal problems.
Inkling also allows users to adjust how much computational effort it applies to a request. A user can choose faster responses for simple tasks or increase the model’s reasoning effort for more complex problems.
This gives organizations greater control over the balance between response quality, speed, and computing costs.
Thinking Machines has acknowledged that Inkling is not the strongest AI model currently available.
Instead of attempting to win every benchmark, the company is presenting Inkling as a balanced and adaptable foundation. Its value is expected to come from what customers can build with it rather than from its performance immediately after release.
This positioning separates Inkling from flagship models promoted as universal assistants capable of handling almost any task.
Thinking Machines believes that a general-purpose model can become more useful when it is trained further on the knowledge, terminology, processes, and preferences of a particular organization.
A financial company, healthcare provider, manufacturer, or software business may require different types of expertise. A single centrally trained model may struggle to represent the specialized knowledge needed across all those industries.
Inkling is closely connected to Tinker, the model customization platform developed by Thinking Machines.
Organizations can use Tinker to fine-tune Inkling using their own data and expertise. This allows them to transform the open model into a more specialized system for internal workflows or customer-facing applications.
Thinking Machines is therefore treating Inkling as a starting point rather than a finished product.
This strategy gives customers more control, but it also places more responsibility on them. Fine-tuning a large AI model requires experienced engineers, reliable training data, careful evaluation, and strong safety controls.
Companies must ensure that their customized versions do not produce harmful, biased, insecure, or inaccurate results. Thinking Machines can provide the tools, but customers will still need the technical talent to use those tools responsibly.
The release of Inkling reflects a larger debate about the future of artificial intelligence.
Companies such as OpenAI, Anthropic, and Google have built popular general-purpose assistants that can perform a wide range of tasks. These systems are convenient because customers can access powerful capabilities without managing the underlying models.
However, closed platforms can give customers less control over how a model is trained, hosted, updated, and customized.
Open-weight models offer a different trade-off. Organizations can operate them in private environments, fine-tune them for specific tasks, and potentially reduce their dependence on an external AI provider.
For businesses handling sensitive information, the ability to keep data and model operations within controlled infrastructure may be particularly attractive.
The strongest general-purpose model may still be preferable for difficult research, advanced reasoning, or experimental work. Specialized open models, however, could handle a large share of routine production tasks more efficiently.
Thinking Machines is emphasizing efficiency rather than absolute model strength.
The company says Inkling can achieve certain coding results while using fewer tokens than some competing open models. Lower token usage can reduce both response times and operating expenses.
This matters because AI costs can grow quickly when a model is deployed across a large organization.
A small performance improvement may be less valuable than a major reduction in computing requirements, especially when the model is completing repetitive tasks thousands or millions of times.
Customization could create additional efficiency gains. A model trained for a narrow business function may require less reasoning and fewer instructions than a general-purpose system that must interpret every request from the beginning.
Thinking Machines argues that much of the world’s most valuable knowledge exists inside organizations and among experienced employees.
A general AI model may understand broad financial concepts, for example, but it may not understand a specific investment firm’s research methods, risk framework, or decision-making process.
By training an open model on internal expertise, a business may be able to create a system that performs better on its specialized tasks than a more powerful general model.
This idea was demonstrated through work involving investment firm Bridgewater Associates. Researchers customized an existing open model using financial knowledge and reported stronger results on financial reasoning evaluations at a lower operating cost.
Such results will require independent testing before broader conclusions can be made. Still, the project illustrates the type of enterprise use case Thinking Machines wants to support.
Releasing model weights can make a technology more accessible, but it can also complicate the path to revenue.
Once organizations download Inkling, they may be able to host and operate it without paying Thinking Machines for every request. This is different from proprietary AI services that charge customers according to usage.
As a result, the company’s commercial strategy is likely to depend heavily on Tinker and related services.
Thinking Machines can generate revenue by helping customers train, customize, evaluate, and host specialized versions of Inkling. It may also earn money through partnerships with infrastructure providers and companies that offer managed access to the model.
The model itself can attract developers and encourage adoption, while the surrounding customization ecosystem becomes the primary commercial product.
Inkling represents a major milestone for a startup that has attracted significant attention since its formation.
Thinking Machines has spent much of its early existence building infrastructure and developing its research direction outside public view. The release of a large open model now gives developers, customers, and competitors something concrete to evaluate.
The company still faces difficult questions about training costs, revenue, adoption, and competition. It must prove that businesses are willing to invest in customizing models rather than simply subscribing to established AI platforms.
It must also demonstrate that open customization can produce reliable advantages without creating unmanageable technical and safety risks.
Even so, Inkling offers a clear statement about the company’s direction. Thinking Machines is betting that the future of enterprise AI will not be controlled entirely by a few universal models.
Instead, it believes organizations will want AI systems shaped by their own expertise, operated under their own control, and optimized for their own work.
Share your thoughts about this article.
Be the first to post a comment!
