Mira Murati’s $12 Billion AI Lab Challenges Big Tech With a Customizable, ‘Good-Enough’ Open Model

• Early build-up: a quiet infrastructure push
• July 15: Inkling is unveiled
• A deliberate step back from the frontier
• Betting on customization and open weights

Mira Murati’s $12 Billion AI Lab Challenges Big Tech With a Customizable, ‘Good-Enough’ Open Model
Mira Murati is re-entering the AI race not by chasing the most powerful frontier system, but by betting that enterprises want a model they can reshape to fit their own needs — even if it’s not the smartest one on the leaderboard.

§Early build-up: a quiet infrastructure push

Over the past year and a half, Thinking Machines Lab has been building AI infrastructure largely out of public view, setting up the groundwork for its own models and a customization platform called Tinker. Backed by a $12 billion valuation, the company is positioning itself as an alternative to renting closed systems from giants like OpenAI, Anthropic, and Google.

§July 15: Inkling is unveiled

On Wednesday, Thinking Machines released its first model, Inkling, an “open-weight” AI system whose full weights can be downloaded, modified, and fine-tuned by external developers and enterprises. Murati described it as the lab’s first foundational model, trained from scratch rather than adapted from another company’s system.

Technically, Inkling is a mixture-of-experts model with 975 billion total parameters, using about 41 billion for any given task, trained on 45 trillion tokens spanning text, image, audio, and video. For now it only outputs text — including code and structured data — but is designed to reason across all four modalities.

§A deliberate step back from the frontier

In its own materials, Thinking Machines stresses that Inkling is not the most performant model available, open or closed, and instead aims for “solid capabilities across the board rather than state-of-the-art performance in a single area.” The company pitches Inkling as a starting point and “first public proof point” of its approach: calibrated answers, tunable “thinking effort,” and enterprise-centric customization over raw benchmark dominance.

§Betting on customization and open weights

Axios frames the launch as a direct bet that enterprises “want AI they can customize rather than simply rent from a handful of frontier labs,” with full weights published on Hugging Face and integrated into Tinker for fine-tuning. Instead of optimizing for general-purpose supremacy, Thinking Machines is prioritizing adaptable models that organizations can host, tailor on proprietary data, and potentially run more cheaply than closed competitors — while already training more powerful successors beyond Inkling’s debut.

Continue reading https://foxvector.com/stories/019f6801-ed60-2adc-7006-242b67611e8a