Mira Murati's Thinking Machines launches Inkling, open-weight AI model

Mira Murati, who departed OpenAI in September 2024, founded Thinking Machines on a platform of accessibility, customization, and multimodal collaboration. On July 15, 2026, the company announced Inkling, the first foundation model fully trained from scratch by the lab. The model is a mixture-of-experts architecture with 975 billion total parameters but activates only about 41 billion parameters per average prompt, enabling faster inference and lower operational costs.

Inkling was trained on approximately 45 trillion tokens spanning text, image, audio, and video. While it can reason natively across all four modalities, output is limited to text, including code, styled artifacts, and structured data. The company released the model as open weights, allowing developers to inspect and modify the full codebase without licensing fees. Additional features include "thinking effort" controls that let users trade processing speed for accuracy, and uncertainty flagging to prevent straightforward hallucinations.

Developers can fine-tune Inkling directly on Tinker, Thinking Machines' training API that launched in October 2025. Early results demonstrated competitive performance: Inkling achieved comparable coding results to Nvidia's Nemotron 3 Ultra model while using two-thirds fewer tokens. In a high-profile collaboration, Bridgewater Associates researchers fine-tuned an open model using Tinker with proprietary financial data, achieving 84.7% accuracy on leading financial reasoning benchmarks—outperforming proprietary alternatives at less than 10% of the cost.

The company trained Inkling on Nvidia's GB300 NVL72 system under a partnership announced in March 2026, completing the full training cycle in less than nine months—far shorter than the multiyear development cycles at OpenAI and Anthropic. Thinking Machines acknowledged that Inkling does not match the most advanced proprietary systems but is betting that customizability and lower total cost of ownership will appeal to organizations. Rather than distribute the model through a rigid chatbot interface, Thinking Machines positions Inkling as a base model for organizations to fine-tune and self-host on their infrastructure. Analysts noted that the business model—monetizing via Tinker rather than API access—could prove more disruptive than the model itself, accelerating the commoditization of large language models and fundamentally shifting how enterprises evaluate and deploy AI systems.

Thinking Machines' launch of Inkling marks a strategic pivot in the open-source AI landscape. The company, which spent the past year primarily announcing funding rounds and its Nvidia partnership, now enters the market with a fully homegrown foundation model trained in under nine months—substantially faster than the multiyear timelines at rivals like OpenAI and Anthropic. The model's open-weights availability directly counters a year-long dominance by Chinese AI firms in the open-source ecosystem, a gap that widened after Meta deprioritized its Llama family in favor of proprietary approaches.

What distinguishes Inkling is not merely its technical specs but Thinking Machines' business model inversion. While OpenAI, Anthropic, and other incumbents monetize through metered API access, Thinking Machines is shifting revenue to Tinker, a platform that enables enterprises to fine-tune and run models on their own infrastructure. This approach converts the AI model itself into a commodity—developers can customize freely without per-token costs—and captures value downstream through tooling and customization. The Bridgewater collaboration exemplifies this: a financial-domain fine-tuned version achieved state-of-the-art performance on specialized benchmarks at a fraction of proprietary model costs, demonstrating that customization economics can offset raw capability gaps.

No comments yet. Be the first to share your thoughts!