Thinking Machines Lab Launches Inkling: Mira Murati's First Open-Weight Model

Thinking Machines Lab — founded by former OpenAI CTO Mira Murati — publicly launched its first AI model on July 15, 2026: Inkling, a 975 billion parameter open-weight Mixture of Experts model trained simultaneously on text, images, audio, and video, and designed specifically for fine-tuning through the company's Tinker platform.

Key Points:

• Inkling is a 975 billion parameter model with 41 billion parameters active during each inference step — making it computationally efficient to run despite its large total size.

• The model was trained natively on text, images, audio, and video simultaneously — making it multimodal from its foundations rather than retrofitted.

• Thinking Machines is positioning Inkling primarily as a fine-tuning foundation — designed to be customized by organizations for specific domains via the Tinker platform, which provides tools and infrastructure to train custom versions on proprietary data.

• The model is open-weight: freely available for download with no API cost, and organizations can run it on their own infrastructure.

• Mira Murati served as OpenAI's CTO from 2018 to September 2024, overseeing GPT-3, GPT-4, DALL-E, Codex, and the launch of ChatGPT. The Thinking Machines research team includes senior former employees from OpenAI, Anthropic, and DeepMind.

Inkling's fine-tuning-first design is strategically distinct from every other major model launch this week: Thinking Machines is not competing on raw benchmark scores, but on how easily organizations can customize the model for their own data and specific use cases.

An open-weight multimodal model at this scale — trained natively on text, images, audio, and video — is unprecedented in the open-weight ecosystem. It brings enterprise-grade multimodal capability to organizations that could not afford or did not want to rely on closed-API multimodal models.

Why It Matters: Inkling represents a direct response to enterprise demand for controllable, customizable AI. Organizations can now fine-tune frontier-class multimodal models on proprietary data without sending that data to external APIs.