Thinking Machines Lab, an AI startup established by former OpenAI CTO Mira Murati, unveiled its first proprietary AI model, Inkling, on Wednesday morning. This model diverges from the major offerings of OpenAI, Anthropic, and Google by being open-weight, allowing external developers and companies to download and modify it freely.
Inkling is a mixture-of-experts system boasting a total of 975 billion parameters, although it employs approximately 41 billion for specific tasks. This design is common among large models as it enhances efficiency and cost-effectiveness. The model was trained on an extensive dataset of 45 trillion tokens encompassing text, images, audio, and video, enabling it to natively reason across all four modalities. Currently, however, Inkling is limited to text outputs, which include code, styled outputs, and structured data.
This launch marks Thinking Machines Lab’s first public demonstration of its AI capabilities after a year and a half of working largely behind the scenes on its infrastructure. Earlier this year, the company shared a preview of its “interaction models,” intended to engage users more dynamically than traditional chatbots. Inkling also tests the core assumption of the startup: that AI models which organizations can adapt themselves will ultimately prove more effective than the generalized models offered by larger labs.
Inkling is designed to provide nuanced responses, alerting users to uncertainty rather than generating random guesses. It also allows users to adjust the “thinking effort” applied during tasks, providing a trade-off between response speed and the depth of analysis required. In comparative benchmarks, the company claims that Inkling operates using a third of the tokens needed by Nvidia’s Nemotron 3 Ultra to achieve similar coding results.
It is important to note that Thinking Machines does not assert that Inkling is the leading model in the market. In its documentation, the company explicitly states that Inkling is "not the strongest model available today, closed or open," and rather aims for balanced performance and adaptability.
The primary target audience for Inkling appears to be enterprises, as Thinking Machines is promoting the model as a foundation for customization rather than a final product. Organizations will be able to refine Inkling through Tinker, the company’s model-customization platform, although this places some responsibility on users to ensure their modifications are secure, which necessitates skilled machine learning knowledge.
Unlike OpenAI’s ChatGPT or Anthropic’s Claude, which focus on general-purpose chatbot functionalities, Inkling is positioned to be a more versatile tool for organizations wishing to tailor their AI solutions.
In a post released by Thinking Machines prior to the launch, the company argued that centrally trained AI models often fail to meet the nuanced needs of specific enterprises, as much of the expertise needed is unique to those organizations. This sentiment echoes growing concern about the value of centralized models, as articulated by Microsoft CEO Satya Nadella, who pointed out that companies effectively incur dual costs when relying on proprietary AI models — initially through subscription fees and subsequently by relinquishing proprietary business insights through usage data.
Hugging Face’s CEO, Clem Delangue, echoed these thoughts, stating during a recent discussion that mainstream AI operations are likely to veer toward private or open-source models as opposed to centralized solutions, aligning with the path that Thinking Machines is pursuing.
Evidence supporting Thinking Machines’ hypothesis emerged from a collaboration with Bridgewater Associates, the largest hedge fund globally (which does not invest in Thinking Machines). The team trained an existing open-source model with Bridgewater's financial expertise, achieving a score of 84.7% on financial reasoning tests that surpassed leading proprietary AI models while incurring significantly lower operating costs — though it’s crucial to mention these results originated from the companies’ own assessments, not independent validation.
Thinking Machines emphasizes its rapid development timeline, claiming to bring Inkling to market in about nine months, substantially quicker than the timelines observed for OpenAI and Anthropic.
Questions have arisen around whether Inkling utilized outputs from competitor models, known as “distillation,” a practice under scrutiny in the industry. According to the company, Inkling was predominantly trained from scratch but did incorporate some early post-training data from other open-weight models, including Moonshot AI’s Kimi K2.5, before transitioning to large-scale reinforcement learning. Future models are anticipated to rely exclusively on self-contained post-training methods.
On the financial front, Thinking Machines has maintained a level of caution. Earlier this year, the company formed a strategic partnership with Nvidia, harnessing substantial computing capacity for training Inkling on Nvidia’s advanced systems. However, the company has been reticent regarding its revenue plans, which reportedly have not been a primary focus. There were discussions about a potential $50 billion fundraising round, but updates indicate those efforts have faced delays.
Another key consideration is the scalability of Thinking Machines’ spending relative to competitors like OpenAI and Anthropic. The startup seems to favor a more efficient spending model, possibly eliminating the necessity to match the financial outlays of larger rivals. The open-weight framework of Inkling means that once the model is in the public domain, there is no obligation for users to pay for its operation, contrasting the subscription models employed by OpenAI and Anthropic. Instead, revenue for Thinking Machines will largely stem from Tinker, focusing on training, fine-tuning, and the surrounding ecosystem.
Thinking Machines’ workforce appears to have stabilized, currently employing around 200 individuals, a recovery from earlier attrition, including the exit of two co-founders who joined OpenAI.
The company cultivates a culture that prioritizes stability over individual prominence, a calculated approach that minimizes disruption when changes occur within the team. This is notably significant as much of the company’s narrative is still linked to its co-founder’s well-known name, balancing recognition with a commitment to organizational continuity.
