Thinking Machines Lab’s ‘Inkling’: A U.S.-Built Open-Weights Multimodal AI Model
Published on: July 18, 2026
In the past 24 to 48 hours, Thinking Machines Lab, a startup led by former OpenAI executives including Mira Murati, unveiled its first foundational model: Inkling. The model distinguishes itself by being openly published—that is, its full weights are available under an Apache‑2.0 license—allowing organizations and developers to download, inspect, and fine‑tune it without relying on private APIs.
Inkling is a sparse mixture‑of‑experts (MoE) model comprising 975 billion total parameters, although it uses only around 41 billion active parameters per token during inference, enabling greater efficiency in deployment. It supports text, audio, and video input natively, without separate encoders, and offers a vast context window of up to one million tokens.
The creators are transparent about its performance: Inkling is explicitly not the strongest model on benchmarks. Instead, its value lies in its adaptability—it’s a multimodal model built from scratch to serve as a customizable foundation that enterprises and researchers can own, fine‑tune, and deploy under their control.
Technically innovative, Inkling introduces architectural features such as a sparse MoE design with dedicated expert routing, efficient multimodal fusion without encoders, and a controllable reasoning‑effort mechanism—enabling dynamic adjustment of computational effort per task.
The release positions Thinking Machines as a credible entrant in the open‑model space, offering an alternative to closed, proprietary models primarily developed by large AI labs or Chinese organizations. Its availability on public platforms like Hugging Face immediately opens it up for experimentation, research, and specialized enterprise use.
No comments yet.