AI startup Thinking Machines Lab has unveiled Inkling, a 975-billion parameter open-weight model designed for enterprise customization. Unlike closed models from OpenAI or Google, this system allows businesses to modify the technology directly, focusing on specialized tasks rather than general-purpose chatbots.
Thinking Machines Lab, an AI startup co-founded by former OpenAI executive Mira Murati, has introduced Inkling, its first proprietary AI model. The company is positioning the system as a direct alternative to the closed-model frameworks currently dominated by industry leaders like OpenAI, Anthropic, and Google. By making Inkling open-weight, the company allows external developers to download and adapt the model, signaling a strategic shift toward customizable, enterprise-focused AI solutions.
Inkling's Technical Design and Scale
Inkling is built using a mixture-of-experts architecture with a total of 975 billion parameters. For individual tasks, it uses roughly 41 billion active parameters, a design intended to increase operational speed and reduce the cost of running the system. The model was trained on 45 trillion tokens of multi-modal data, including text, audio, images, and video. A key feature of the model is its ability to flag uncertainty in its outputs, which is meant to provide more reliable answers for business applications compared to models that generate responses without expressing confidence levels. Users can also scale the amount of computing power applied to a task to balance speed and accuracy.
Enterprise Strategy and Revenue Model
Unlike competitors that prioritize general-purpose chatbots for individual users, Thinking Machines is targeting the enterprise market. The company argues that organizations often lose specialized knowledge when relying on standardized models. To address this, it launched a platform called Tinker, which companies can use to fine-tune Inkling based on their own internal expertise. The company’s revenue strategy focuses on this customization platform, as well as managed hosting services, rather than charging subscription fees for direct model access. Thinking Machines reported reaching a revenue-generating stage within nine months of development, a significantly faster timeline than many of its established peers.
Industry Shift Toward Open-Weight AI
This launch reflects a broader industry debate regarding the control and cost of AI models for businesses. Microsoft CEO Satya Nadella has previously noted that enterprises using closed proprietary models may face hidden costs, such as the potential loss of proprietary business data. Similarly, Hugging Face CEO Clem Delangue has suggested that more production-level AI work will move toward open or private models, reserving large frontier models for highly specialized functions. Thinking Machines has secured a strategic investment from Nvidia and utilized Nvidia’s GB300 NVL72 systems to train its new model. Investors and developers will likely monitor how effectively the company can compete for enterprise contracts against larger, more established labs as it balances the costs of high-performance computing with the demand for specialized, private AI infrastructure.