Ex-OpenAI CTO’s $12B Startup Launches ‘Good-Enough’ Open Model to Undercut Frontier AI Giants

• Early build-up: money, mission, and a different bet
• July 15: Inkling is introduced
• Customization over raw power
• Community and competitive context

Ex-OpenAI CTO’s $12B Startup Launches ‘Good-Enough’ Open Model to Undercut Frontier AI Giants
Thinking Machines Lab is betting that a highly customizable, “good-enough” open model can matter more to businesses than competing in the frontier AI arms race.

§Early build-up: money, mission, and a different bet

After leaving OpenAI, former CTO Mira Murati raised about $2 billion for Thinking Machines at a $12 billion valuation to build an alternative to one-size-fits-all AI offered by giants like OpenAI, Anthropic, and Google. The company spent roughly a year and a half building infrastructure largely out of public view before revealing its strategy: open-weight models that enterprises can adapt and control themselves, rather than renting closed systems.

§July 15: Inkling is introduced

On Wednesday, Thinking Machines released its first model, Inkling, describing it as a foundational, open-weight system with solid but not best-in-class performance. Murati announced that it was “trained from scratch” with open weights, fine-tunable on the company’s customization platform, Tinker. The lab stresses that Inkling “is not the most performant model available today, closed or open,” but is meant as a base for future models.

TechCrunch detailed the technical design: a mixture-of-experts architecture with about 1 trillion parameters, 41 billion active per task, trained on trillions of multimodal tokens and intended as a test of the startup’s core bet that adaptable AI will beat generic tools in the enterprise.

§Customization over raw power

Axios framed Inkling as a direct challenge to the dominant rental model: enterprises “want AI they can customize rather than simply rent” from a handful of labs, even if that means using a model that is not the strongest overall. The model’s full weights are available on platforms like Hugging Face and live for fine-tuning on Tinker, aligning with growing demand for open-weight systems companies can deploy on their own infrastructure.

§Community and competitive context

Online, early reactions from the open-source and tooling ecosystem focused on scale and openness. One widely shared post called Inkling “first ever open and large (1T), text, image and audio in, text out,” highlighting how quickly developers could accelerate it with minimal code changes. Another emphasized it as a foundation model “with solid performance across a broad categories of capabilities, for use in practice and customization.”

At the same time, coverage noted that Inkling draws on data generated by existing open models, including Chinese lab Moonshot AI’s Kimi K2.5, reflecting an increasingly interdependent — and geopolitical — AI training ecosystem. As Inkling climbed trending charts on Hugging Face, the debut marked Thinking Machines’ first concrete step toward its larger goal: proving that open, adaptable AI can rival frontier giants in real-world use, even if it doesn’t top every benchmark.

Continue reading https://foxvector.com/stories/019f6a94-f74d-24fa-7338-0363a5a5e70a