Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has officially launched Inkling, its first proprietary open-weight large language model. Released Wednesday, Inkling represents a decisive pivot toward the company’s core thesis: that AI architectures capable of deep organizational customization will ultimately outperform centralized, one-size-fits-all proprietary models. Unlike dominant industry offerings from OpenAI, Anthropic, and Google, Inkling’s weights are publicly accessible, enabling developers and enterprises to download, modify, and fine-tune the system without vendor lock-in. Technically, Inkling utilizes a mixture-of-experts architecture with 975 billion total parameters, activating approximately 41 billion parameters per inference to optimize speed and operational costs. The model was trained on 45 trillion tokens spanning text, images, audio, and video, granting it native multimodal reasoning capabilities. Thinking Machines explicitly positions Inkling as a well-rounded foundation rather than a best-in-class benchmark contender. The system features calibrated response mechanisms that flag uncertainty and allows users to adjust computational effort dynamically. Preliminary evaluations suggest Inkling achieves comparable coding performance to Nvidia’s Nemotron 3 Ultra while consuming significantly fewer inference tokens. The release underscores a broader industry tension regarding AI deployment strategies. Thinking Machines markets Inkling less as a finished consumer product and more as an enterprise starting point, integrated with its Tinker customization platform. This approach aligns with growing executive skepticism toward proprietary models. Recent statements from Microsoft CEO Satya Nadella and Hugging Face CEO Clem Delaigue emphasize the economic and strategic drawbacks of centralized AI, predicting a shift toward private or open-source alternatives for production workloads. Thinking Machines points to a collaboration with Bridgewater Associates, where fine-tuning an open model on proprietary financial data reportedly achieved top-tier reasoning scores at a fraction of traditional cloud inference costs. Behind the launch, the company has accelerated its development timeline, claiming a nine-month turnaround from inception to public release, significantly faster than competitors like OpenAI and Anthropic. Training was conducted entirely on Nvidia’s GB300 NVL72 systems, supported by a strategic partnership securing gigawatt-scale computing capacity. While the company acknowledges partial use of other open-weight models for early post-training data generation, it commits to fully self-contained training pipelines for future iterations. Financially, Thinking Machines maintains a lean operational posture, currently employing approximately 200 staff members and steering clear of heavy marketing campaigns. The business model explicitly relies on Tinker’s fine-tuning services and a hosted ecosystem rather than direct model licensing or API metering. This efficiency-first strategy suggests the startup aims to compete through economic sustainability and architectural adaptability rather than capital-intensive benchmark scaling. With Inkling now available, Thinking Machines is positioning itself at the forefront of the open-weight enterprise AI movement, betting that organizational control and customization will define the next phase of industrial artificial intelligence.