Mira Murati's Thinking Machines Launches Inkling, an Open-Weight Multimodal Model

1. Mira Murati's Thinking Machines Launches Inkling, an Open-Weight Multimodal Model

On July 15, 2026, Thinking Machines released Inkling, its first model trained from scratch. The weights are fully open and available immediately for fine-tuning on Tinker, the company's own platform. Inkling reasons across text, image, and audio modalities. The announcement came from Mira Murati, former CTO of OpenAI and Thinking Machines' founder, and the post pulled 1.2 million views within roughly 24 hours. A public Inkling Playground is also live for direct experimentation.

The release lands Thinking Machines squarely in the competitive space Meta has been building since Llama 2: open weights, fine-tunable by anyone, multimodal. The difference is the founder signal. Murati carries institutional credibility that most open-source labs cannot buy, and pairing open weights with a proprietary fine-tuning platform on day one is a deliberate play. It mirrors the Hugging Face model: give away the weights, monetize the tooling. If Tinker gains traction as the preferred fine-tuning surface for Inkling derivatives, Thinking Machines controls the workflow layer even as the model itself is free. That is a structurally different business than selling API access, and it puts pressure on Mistral and Cohere, both of whom charge for fine-tuning infrastructure.

The pattern worth watching here is whether high-profile post-OpenAI founders converge on open weights as a go-to-market wedge rather than a philosophical stance. Ilya Sutskever's SSI has signaled the opposite direction, staying closed. Murati's choice draws a clear line. The next signal to track: whether enterprise customers fine-tune on Tinker at scale, or pull the weights and self-host, which would test whether the platform bet holds.

Source: @miramurati on X