New AI Prioritizes Customization Over "One-Size-Fits-All"

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has launched Inkling, its inaugural open-weight AI model. This means outside developers and companies can freely download and customize the system.

Inkling is engineered as a mixture-of-experts system, boasting 975 billion parameters but utilizing around 41 billion for specific tasks to optimize speed and cost. It was trained on 45 trillion tokens across text, image, and audio, natively reasoning across these modalities.

The company's core strategy is that AI which organizations can adapt will outperform standardized models. Inkling offers calibrated answers, flags uncertainty, and allows adjustable "thinking effort" for speed-accuracy trade-offs.

This model is positioned as a starting point for enterprises to fine-tune on Thinking Machines' Tinker platform, contrasting with general-purpose chatbots from OpenAI, Anthropic, and Google.

The argument for customizable AI is gaining traction, with Microsoft CEO Satya Nadella noting enterprises pay double for proprietary models. Hugging Face CEO Clem Delangue also predicts a shift towards private or open-source alternatives for production AI work.

A project with Bridgewater Associates demonstrated an adapted open-source model outperforming top proprietary AIs on financial reasoning tests at a fraction of the cost.

Thinking Machines highlights rapid development, achieving market presence and revenue in approximately nine months, significantly faster than competitors like OpenAI and Anthropic.

While Inkling used some data from other open-weight models during pretraining, the company plans fully self-contained post-training for future iterations. Revenue generation is expected from the Tinker platform for training, fine-tuning, and hosting services.