Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI CTO Mira Murati, has released its first in-house AI model. Called Inkling, the model is an open-weight system that allows outside developers and companies to download and modify it directly—marking an important step in the startup’s effort to challenge the one-size-fits-all approach taken by major AI labs.
 |
| Credit: Patrick T. Fallon/AFP / Getty Images |
Inkling is a 975-billion-parameter mixture-of-experts model
Inkling uses a mixture-of-experts architecture with 975 billion total parameters. However, it activates only about 41 billion parameters for a given task, a design approach intended to make very large AI models faster and less expensive to run.
According to Thinking Machines Lab’s release materials, the model was trained on 45 trillion tokens covering text, images, audio and video. It also reasons natively across all four types of data.
For now, however, Inkling’s outputs are limited to text. The model can generate code, styled artifacts and structured data, but the company’s description of its training and reasoning capabilities extends across multiple forms of information.
The first public proof point for Thinking Machines Lab
Inkling is the first major public proof point from Thinking Machines Lab after roughly a year and a half of building AI infrastructure largely out of public view.
Some of the startup’s work became visible earlier through a May research preview focused on “interaction models.” These systems were designed to listen and speak, and could even interrupt, rather than following the typical chatbot pattern of stopping and waiting for a user to respond.
Inkling now offers a clearer look at the broader strategy behind the company. Thinking Machines Lab is betting that AI models organizations can adapt for themselves may ultimately outperform the one-size-fits-all systems currently offered by the biggest AI labs.
Inkling can adjust its “thinking effort”
One of Inkling’s key features is its focus on calibrated answers. The model is designed to flag uncertainty instead of simply guessing, according to the company.
Users can also adjust the model’s “thinking effort” depending on whether they want to prioritize deeper reasoning or faster responses. This gives users a way to trade off performance and speed based on the demands of a particular task.
The company also highlights Inkling’s efficiency on coding workloads. On one benchmark, Thinking Machines says Inkling uses one-third as many tokens as Nvidia’s Nemotron 3 Ultra—described as the company’s latest-generation open-weight model—while achieving the same coding performance.
Thinking Machines is not claiming Inkling is the strongest AI model
The company is taking a notably measured position on Inkling’s overall capabilities.
Thinking Machines Lab’s newest blog post explicitly states that Inkling is “not the strongest overall model available today, open or closed.”
Instead, the model appears to be aimed at delivering well-rounded performance while giving organizations the ability to adapt the system to their own requirements.
That positioning separates Inkling from a straightforward race to claim the highest benchmark scores. The company’s broader argument is that flexibility and customization could be just as important as raw model performance for organizations building AI systems around specific workflows.
An open-weight model aimed at customization
The open-weight nature of Inkling is central to its launch. Unlike flagship models from OpenAI, Anthropic or Google, outside developers and companies can download the model and modify it directly.
This approach is particularly relevant to organizations that want to adapt an AI model instead of using a general-purpose system in its original form.
Thinking Machines Lab is currently presenting Inkling less as a finished, ready-made product and more as a starting point for organizations that want to fine-tune it for their own needs. The company’s strategy is connected to Tinker, its model-customization platform, which is intended to support organizations in adapting AI systems themselves.
That approach reflects the startup’s broader thesis: organizations may gain more value from AI systems that can be shaped around their own requirements than from a single model designed to serve everyone in the same way.
Inkling puts Thinking Machines Lab’s strategy to the test
The release represents more than the launch of a new AI model. It is also an early test of Thinking Machines Lab’s central bet on customizable artificial intelligence.
The company has spent approximately a year and a half building its infrastructure, with much of that work taking place away from the public spotlight. Inkling now gives developers and organizations a tangible model through which to evaluate the startup’s approach.
Its combination of open weights, a large mixture-of-experts architecture, multimodal training and adjustable thinking effort is intended to support a range of uses while leaving room for further customization.
At the same time, Thinking Machines Lab is not presenting Inkling as the overall leader among open or closed AI models. Its stated focus is instead on creating a flexible foundation that organizations can adapt to their own requirements.
What Inkling means for the company’s next phase
Inkling gives Thinking Machines Lab its first public in-house model and provides a clearer view of how the startup intends to compete in an increasingly crowded AI market.
Rather than positioning the model solely around being the biggest or strongest system, the company is emphasizing adaptability, efficiency and the ability for organizations to shape the model themselves.
The launch also builds on the startup’s earlier work in interaction models, showing a broader effort to develop AI infrastructure and systems that can be customized for different users and organizations.
For now, Inkling’s biggest significance may lie in the strategy it represents. Thinking Machines Lab is betting that the future of AI will not be defined only by increasingly powerful general-purpose models, but also by systems that organizations can modify and tailor to their own needs.
Thinking Machines Lab has released Inkling as its first in-house open-weight AI model, combining a 975-billion-parameter mixture-of-experts architecture with adjustable thinking effort and a focus on customization. The company says Inkling is not the strongest overall model available, but its goal is to provide a well-rounded foundation that organizations can adapt themselves.
With Inkling now public, developers and companies can begin evaluating the model as a potential starting point for customized AI systems through the company’s broader model-adaptation approach.
Explore the Inkling release and follow Thinking Machines Lab’s next developments as the company continues building its vision for adaptable AI.
إرسال تعليق