Former OpenAI CTO Mira Murati’s startup, Thinking Machines, just launched Inkling- an open-weight MoE model built to challenge one-size-fits-all AI.
If you thought the AI race was purely about building one massive, closed-door chatbot to rule them all, Mira Murati’s new venture just flipped the script. Thinking Machines Lab has officially launched Inkling, its first open-weights model, throwing a brilliant wrench into the one-size-fits-all AI narrative.
Inkling is a powerhouse of 975 billion parameters that processes text, images, and audio natively. But the real headline is the philosophy behind it.
Thinking Machines is releasing the weights under an enterprise-friendly Apache 2.0 license, which perfectly pairs with their fine-tuning ecosystem, Tinker. This is instead of boxing developers into a rigid, subscription-style sandbox.
The strategic nuance here is commendable.
To ship a model of this magnitude in just nine months, Murati’s team leaned into data distillation, using footprints from existing open models such as Moonshot AI’s Kimi K2.5 to bootstrap training. This is a masterclass in modern engineering efficiency. They realized that designing a practically adaptive foundation rapidly matters far more than waiting years for an isolated system from scratch.
This is a massive win for open tech.
By actively betting against monolithic, centralized AI architectures, Thinking Machines is proving that the future belongs to specialization.
Inkling isn’t trying to be a singular, omniscient oracle for the entire planet. It is designed to be a deeply customizable framework that engineers can actually sculpt to fit their unique business logic.
By putting the weights directly in the hands of creators, they remind us that the best AI isn’t one we eventually build upon.
