Mira Murati is using her $12 billion AI startup, Thinking Machines Lab, to bet that “good-enough” and fully customizable AI will beat the smartest one-size-fits-all systems sold by today’s biggest labs.

On July 15, Thinking Machines unveiled Inkling, its first in-house AI model, built largely out of public view over the past 18 months.¹ Unlike the closed flagships from OpenAI, Anthropic, and Google, Inkling’s weights are open, meaning companies can download and directly modify the system.² Murati had already framed the release on X, calling Inkling the company’s “first model… trained from scratch, weights are open, fine-tunable on Tinker today.”clem 🤗 (ClementDelangue)³

Early technical details trickled out through the launch day. TechCrunch reported that Inkling is a mixture‑of‑experts model with about 975 billion parameters, of which roughly 41 billion are active per query, trained on 45 trillion tokens spanning text, image, audio, and video, but currently outputting only text.² Axios emphasized that Thinking Machines is “clear that Inkling is not the strongest model available,” instead pitching it as a customizable foundation that can yield better performance and lower costs in specific enterprise settings.⁴

Internally, the lab reinforced that strategy. Research leader Lilian Weng described Inkling as “our open weights model” that aims to provide “solid performance across a broad categories of capabilities, for use in practice and customization,” inviting users to “play it on Tinker.”Lilian Weng (lilianweng)⁵ The company positioned Tinker as the fine‑tuning and deployment hub, and released full weights on platforms such as Hugging Face, where Inkling quickly appeared among trending models.⁴clem 🤗 (ClementDelangue)⁶

Externally, the launch fed into wider industry debates. Axios framed Murati’s move as a direct challenge to expensive, opaque closed models that leave enterprises “renting” intelligence rather than owning it.⁴ The Financial Times, noting Thinking Machines’ $2 billion raise at a $12 billion valuation, highlighted that Inkling’s training drew on data from rival open models, including Chinese lab Moonshot AI’s Kimi K2.5, underscoring intensifying cross‑border competition in AI foundations.⁷

As benchmark races continue, Murati’s first public model is less a finished product than a test of a thesis: that adaptable, open‑weight AI tailored in‑house will ultimately beat frontier giants at their own game.