Thinking Machines Releases Inkling, Its First Open-Weight Multimodal Model Under Apache 2.0

My take: That Mira Murati, former CTO of OpenAI, chose to launch her first model under the Apache 2.0 license says something important: open-weight models with commercial use rights are genuinely competing with closed ones. Inkling has 975 billion total parameters, works as a native multimodal system (text, image, audio, and video), and any company can download it, fine-tune it on their data, and deploy it without paying per query.

For those of us who build with AI, that has practical implications: the difference between depending on a third-party API and having control over the model you use is not just about cost, but about privacy, flexibility, and long-term continuity. A model you can fine-tune on your own data, run on your own infrastructure, and adapt to your industry has value that per-token pricing does not capture.

The design philosophy is also worth noting: instead of claiming the top spot on every benchmark, Thinking Machines prioritized calibrated answers. The model acknowledges uncertainty rather than guessing. That is not a limitation, it is what you need when the margin for error has real consequences.

Does your company have AI use cases where a customizable model under your own control would make more sense than paying for an API month to month?