Thinking Machines, releases first open-weight AI model Inkling
Original: Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling
Why This Matters
Open-weight enterprise AI customization is emerging as a direct strategic challenge to centralized AI lab models.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first open-weight AI model Inkling on July 15. The mixture-of-experts model has 975B total parameters, activates ~41B per task, and was trained on 45 trillion multimodal tokens.
Thinking Machines Lab released Inkling, its first proprietary AI model, on July 15, 2026. Unlike flagship models from OpenAI, Anthropic, or Google, Inkling is open-weight — meaning developers and organizations can download and modify it directly.
Inkling is a mixture-of-experts (MoE) architecture with 975 billion total parameters, activating approximately 41 billion for any given task. It was trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all modalities.
The company explicitly states that Inkling is "not the strongest model available today, closed or open," positioning it instead as a well-rounded starting point for enterprise customization. On one benchmark, Thinking Machines claims Inkling uses one-third as many tokens as Nvidia's Nemotron 3 Ultra to achieve equivalent coding performance. Key design features include calibrated uncertainty flagging and an adjustable "thinking effort" dial to trade accuracy for speed.
Inkling is intended to be used alongside Tinker, the company's model-customization platform, allowing enterprises to fine-tune the model for domain-specific needs. This release marks Thinking Machines' first public milestone after roughly 18 months of development, following a May 2026 research preview of "interaction models" — AI capable of real-time listening, speaking, and interrupting.