Thinking Machines Lab, the innovative AI startup founded by former OpenAI CTO Mira Murati, has officially unveiled its first in-house artificial intelligence model: Inkling. This release marks a significant departure from the industry giants, as Inkling is an open-weight model, empowering external developers and companies to download and directly modify it for their specific needs. It’s a bold move that signals the company’s strong belief in adaptable, custom AI over a one-size-fits-all approach.
After a year and a half of quiet development, Inkling serves as Thinking Machines Labs’ inaugural public demonstration of its AI infrastructure. This model is a direct challenge to the prevailing trend of proprietary, general-purpose models offered by major labs, advocating for AI that organizations can truly make their own.
Inkling: A New Blueprint for Enterprise AI
Inkling is built on a sophisticated mixture-of-experts (MoE) system, boasting 975 billion total parameters. However, it intelligently activates only a fraction—around 41 billion—for any given task, a design choice that ensures speed and cost-efficiency for such a massive model. This efficiency is critical for practical enterprise deployment.
Trained on an expansive 45 trillion tokens of diverse data encompassing text, image, audio, and video, Inkling is designed to reason natively across all four modalities. While its foundational capabilities are multi-modal, current outputs are focused on text, including code, stylized content, and structured data, offering immediate utility for a wide range of business applications.
Thinking Machines doesn’t claim Inkling is the absolute best-in-class model available today, acknowledging that its strength lies elsewhere. Instead, it’s positioned as a highly efficient and well-rounded performer, designed specifically to be a powerful starting point. For instance, the company highlights Inkling’s coding performance, stating it uses a third as many tokens as Nvidia’s Nemotron 3 Ultra to achieve comparable results.
Inkling also offers unique features for precision and control. It delivers calibrated answers, proactively flagging uncertainty rather than making unsupported guesses, and allows users to adjust “thinking effort” to balance speed against thoroughness. This flexibility is key to its utility as a foundational tool for enterprise customization.
The Growing Case for Adaptable AI
Thinking Machines’ core philosophy is that AI trained and controlled centrally will inevitably underperform models tailored by organizations themselves. This is because deeply specialized knowledge often resides within the people and processes of individual companies, which a generic model cannot fully capture. Inkling is therefore marketed less as a finished product and more as a powerful base for organizations to fine-tune using Tinker, the company’s dedicated model-customization platform.
This vision resonates with growing sentiment within the broader tech industry. Microsoft CEO Satya Nadella, for example, has warned that enterprises relying on proprietary AI models effectively “pay twice”—once for subscription costs, and again by ceding valuable business knowledge embedded in their prompts and corrections. This knowledge can then be absorbed into future versions of the vendor’s model, creating a subtle but significant vendor lock-in.
Hugging Face CEO Clem Delangue shares a similar outlook, predicting that advanced “frontier models” will increasingly be reserved for high-value research and experimentation. Meanwhile, most practical production AI work will migrate towards private or open-source alternatives, precisely the market segment Thinking Machines is built to serve. This shift underscores a broader industry move towards greater control and ownership of AI capabilities.
A compelling demonstration of this approach comes from a project with Bridgewater Associates, the world’s largest hedge fund. Researchers from both companies fine-tuned an existing open-source model with Bridgewater’s proprietary financial expertise. The customized model subsequently achieved an impressive 84.7% on financial reasoning tests, surpassing leading proprietary AI models while costing approximately one-fourteenth as much to run. While these results are from a joint evaluation, they strongly support Thinking Machines’ argument for specialized, adaptable AI.
Speed, Strategy, and the Future of AI Models
Thinking Machines has moved with remarkable speed, bringing its technology to market and demonstrating revenue potential in roughly nine months. This rapid timeline contrasts sharply with the longer development cycles of industry predecessors, such as OpenAI (around five years) and Anthropic (approximately three years), highlighting an agile and focused development strategy.
Regarding its training data, Inkling was largely pre-trained from scratch. However, to accelerate early post-training data generation before large-scale reinforcement learning took over, Thinking Machines did utilize outputs from other open-weight models, including Moonshot AI’s Kimi K2.5. The company has stated its commitment to using fully self-contained post-training methods for future models.
Financially, Thinking Machines has remained somewhat guarded about its costs, despite a significant partnership with Nvidia in March. This collaboration grants access to a gigawatt of Vera Rubin computing capacity, with Inkling itself trained entirely on Nvidia’s GB300 NVL72 systems. Given that Inkling is open-weight, the company’s revenue model shifts away from metered access fees, instead relying on its Tinker platform for revenue generation through training, fine-tuning services, and a share of the hosting ecosystem that emerges around its models.
With approximately 200 employees, Thinking Machines has solidified its team, even after some notable departures earlier this year. The company emphasizes a culture of continuity over reliance on individual personalities, a deliberate choice that makes the organization more resilient and adaptable. This foundational stability ensures that the focus remains squarely on delivering innovative, customizable AI solutions for the enterprise.
Source: TechCrunch – AI
