Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first in-house AI model called Inkling on Wednesday morning. It’s also open-weight, unlike flagship models from OpenAI, Anthropic, and Google. This means that external developers and companies can download and modify it directly.
Inkling is an expert-mixed system with a total of 975 billion parameters, but only a fraction (about 41 billion) are used for specific tasks, resulting in a common design that allows very large models to run faster and cheaper. It was natively trained on 45 trillion tokens of text, images, audio, and video, and reasons for all four, according to company release materials. But for now, its output is limited to text, including code, styled artifacts, and structured data.
The model is Thinking Machines Labs’ first public proof point after spending a year and a half building its AI infrastructure mostly in public. Some of that work was already evident in the May research preview of “Interaction Models.” It’s an AI designed to listen and speak (and even interrupt) rather than stand and wait like a typical chatbot. It’s also a test of the startup’s central bet: that AI that organizations can adapt themselves will outperform the generic models currently being sold by the biggest labs.
Inkling is designed to provide tailored answers, such as flagging uncertainties rather than guesses, and allows users to increase or decrease their “thought effort” in exchange for speed. In one benchmark, Inkling uses one-third as many tokens as Nvidia’s Nemotron 3 Ultra (the latest generation open-weight model) to achieve the same coding performance, according to the company.
Thinking Machines does not claim that Inkling is best in class. In its latest blog post, it specifically states that Inkling is “not the most powerful overall model currently available, open or closed.” What it clearly aims for instead is balanced performance.
This begs the question: Who is this product really aimed at within the targeted enterprise market? For now, Thinking Machines is selling Inkling not as a finished product but as a starting point, something that organizations can tweak through Tinker, the company’s model customization platform. This also means, for example, that the customer, not Thinking Machines, is responsible for ensuring that their customizations are secure. (Fine adjustments require serious mechanical talent.)
OpenAI, Anthropic, and Google have taken very different approaches with ChatGPT, Claude, and Gemini, respectively. All of these are first built to compete as general-purpose chatbots, with agent and autonomous capabilities layered on top.
A post published by Thinking Machines last week was clearly intended as a backdrop for this release. In its post, the company argued that AI that is intensively trained and then entrenched by one company will be less powerful than AI that organizations form themselves because much of the expertise is unique to the people who possess it.
Other arguments against closed models are also gaining traction. In a blog post published Sunday, Microsoft CEO Satya Nadella (whose company has invested billions of dollars in both OpenAI and Anthropic) warned that companies using proprietary AI models will effectively be paying twice. The first is a subscription fee, and the second is that business knowledge embedded in prompts and fixes can be carried forward and absorbed into future model versions.
HugFace CEO Clem DeLang made a similar prediction in a conversation with TechCrunch last week. He says frontier models will increasingly be reserved for experimentation and high-value tasks, while most production AI work will move to private or open source alternatives, such as the precise partitioned models that Thinking Machines is building.
The clearest argument for Thinking Machines’ approach comes from a recent project with Bridgewater Associates, the world’s largest hedge fund (which is not originally an investor in Thinking Machines). Researchers from both companies took existing open source models and further trained them based on Bridgewater’s unique financial expertise. The results achieved a score of 84.7% on a financial reasoning test, outperforming top proprietary AI models and reportedly cost approximately 14 times less to run. However, these results are based on the companies’ own evaluations, not independent evaluations.
Either way, Thinking Machines emphasizes how quickly they got here. OpenAI took about five years to bring its technology to market and show profitability, while Anthropic took about three years. Thinking Machines says it achieved similar results in about nine months.
One might wonder if Inkling was trained based on output from a competitor’s model. This is known as “distillation,” which has drawn scrutiny across the industry. According to the company’s own materials, the short answer is part of the answer. Thinking Machines pre-trained Inkling from scratch, but said it used other open-weight models, including Moonshot AI’s Kim K2.5, to generate some of the initial post-training data before large-scale reinforcement learning took over. The company claims that its next model will instead use fully self-contained post-training.
On the cost side, Thinking Machines is more cautious. The company partnered with Nvidia in March to deploy gigawatts of Vera Rubin computing power and train Inkling entirely on Nvidia’s GB300 NVL72 systems, but the company hasn’t said how it plans to cover those costs, and in the view of many, revenue wasn’t a priority. (According to reports, a $50 billion funding round was said to be closing last November, but stalled by January. The company has since declined to discuss the status of the funding.)
A related question is whether Thinking Machines’ spending will ever reach the scale of OpenAI or Anthropic, or whether its efficiency-focused approach means a change in economics. Put another way, the company’s bet may be less that it won’t end up having to spend as much as its larger rivals, and less that it won’t have to. That’s because once the weights are made public, there is no obligation for those who download them to pay Thinking Machines to run them, unlike the pay-as-you-go access sold by OpenAI and Anthropic. It’s not the model itself, it’s Tinker, and the company’s revenue should come through training, tweaking, and now part of the hosting ecosystem built around it.
At least employee numbers seem more stable. Thinking Machines now has about 200 employees, an increase from the level reported after a spate of departures earlier this year, including two co-founders who left for OpenAI in January.
Thinking Machines doesn’t seem interested in overhyping individual movements, as much of the industry does. Company sources say the company’s culture, by design, emphasizes continuity rather than relying on specific personalities. That’s natural. People who aren’t put on a pedestal from the get-go are less likely to suffer setbacks when changing teams. It’s also remarkable that the company would insist on it, given that so much of its company’s story is still tied to its now-famous co-founder’s name, whether she planned it or not.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
