Thinking Machines Lab, spearheaded by former OpenAI CTO Mira Murati, has introduced Inkling, their inaugural in-house artificial intelligence model. This open-weight system represents a strategic pivot against the prevailing "one-size-fits-all" philosophy championed by industry giants like OpenAI, Anthropic, and Google. Inkling empowers external developers and organizations to directly download and modify its architecture, fostering a new era of adaptable AI solutions. Its debut marks a significant milestone after a period of intensive, largely unpublicized, AI infrastructure development.
Inkling is built upon a sophisticated mixture-of-experts architecture, incorporating a colossal 975 billion parameters. However, for any specific task, it intelligently activates only a fraction—approximately 41 billion—of these parameters. This design choice is critical for maintaining computational efficiency and cost-effectiveness, particularly for very large models. The model underwent extensive training on a vast dataset comprising 45 trillion tokens, spanning diverse modalities such as text, imagery, audio, and video, demonstrating a native capacity for reasoning across all four. Despite its multimodal training, Inkling's initial outputs are focused on text generation, including programming code, styled content, and structured data formats.
The release of Inkling serves as Thinking Machines Labs' initial public demonstration of its capabilities, following eighteen months of discreet work on its AI foundation. While some of their prior endeavors, such as "interaction models" designed for more dynamic, conversational AI experiences, had previously emerged in a May research preview, Inkling provides a concrete validation of the startup's core hypothesis: that AI solutions tailored to an organization's unique needs will ultimately outperform generic models offered by leading AI laboratories. This emphasis on customization allows for more precise and relevant AI applications, diverging from the broader utility focus of competitor products.
A key feature of Inkling is its design for calibrated responses, which includes the ability to signal uncertainty rather than making unsupported inferences. Users can also fine-tune the model's "thinking effort," balancing speed against accuracy as required. Thinking Machines claims that Inkling achieves comparable coding performance to Nvidia's Nemotron 3 Ultra, a leading open-weight model, while utilizing only one-third of the tokens. This efficiency underscores the company's commitment to optimized performance and resource utilization.
Thinking Machines readily acknowledges that Inkling may not be the most potent AI model currently available, whether proprietary or open-source. Instead, their objective is to deliver a solution characterized by well-rounded performance and extensive customizability. This approach raises questions regarding its target audience within the enterprise sector, as Inkling is positioned more as a foundation than a final product. It is intended for organizations to further refine through Tinker, the company's dedicated model customization platform, implying that customers will assume responsibility for ensuring the safety and efficacy of their tailored AI systems. This strategy necessitates a high level of machine learning expertise within client organizations.
The company's philosophy gained traction, echoed by Microsoft CEO Satya Nadella, who cautioned against the hidden costs of proprietary AI models. Nadella highlighted that enterprises using such models incur expenses not only through subscriptions but also by unknowingly contributing their valuable business intelligence to enhance future model iterations. Similarly, Hugging Face CEO Clem Delangue posited that frontier models would increasingly be reserved for niche experimentation, with mainstream AI production shifting towards private or open-source alternatives, aligning perfectly with Thinking Machines' market strategy.
A collaborative project with Bridgewater Associates, the world's largest hedge fund, provided compelling evidence for Thinking Machines' argument. Researchers from both entities refined an existing open-source model using Bridgewater's proprietary financial expertise. This customized model reportedly achieved an 84.7% accuracy on financial reasoning tests, surpassing top proprietary AI models, and operated at approximately one-fourteenth of the cost. Although these results were internally validated, they underscore the potential of tailored AI solutions.
One notable aspect is Thinking Machines' accelerated development timeline. The company achieved its current market presence in approximately nine months, a significantly shorter period compared to OpenAI's five years or Anthropic's three years to bring their technologies to market and generate revenue. This rapid progression highlights the team's efficiency and focus in a competitive landscape.
The question of Inkling's training data, particularly concerning the use of outputs from competitor models (a practice known as "distillation"), has been addressed by Thinking Machines. While the model was primarily pre-trained from scratch, the company disclosed that it utilized other open-weight models, including Moonshot AI's Kimi K2.5, to assist in generating early post-training data before large-scale reinforcement learning took over. Thinking Machines has committed to employing fully self-contained post-training methods for its subsequent models, indicating a move towards greater independence in its data sourcing.
On the financial front, Thinking Machines has been less transparent regarding its expenditures. Despite securing a partnership with Nvidia in March to utilize a gigawatt of Vera Rubin computing capacity for Inkling's training on Nvidia's GB300 NVL72 systems, the company has not revealed its funding mechanisms or how it plans to cover these substantial costs. Reports of a potential $50 billion fundraising round in November stalled by January, and the company has since refrained from discussing its financial standing. The company's business model relies on Tinker, its customization platform, for revenue generation through training, fine-tuning, and a share of the hosting ecosystem, rather than direct sales of the open-weight Inkling model.
The company's headcount has stabilized, with approximately 200 employees, despite earlier departures, including two co-founders who joined OpenAI in January. Thinking Machines cultivates a culture that prioritizes continuity over individual reliance, a unique stance in an industry often driven by prominent personalities. This approach suggests that changes in personnel have less impact when the company's ethos emphasizes collective strength and a systemic approach to innovation, even as its public image remains strongly linked to its renowned co-founder.
\
