New US-made AI Model Designed for Customization

Thinking Machines, founded by Mira Murati, launched Inkling on July 15, 2026, as its first model trained from scratch. The model is available on Hugging Face under the Apache 2.0 license, allowing anyone to download and retrain it on their own data.

Inkling's architecture adopts the Mixture of Experts (MoE) approach. It features 66 layers with 256 experts per layer, but only 6 are activated per token. This design ensures that out of 97.5 billion total parameters, only 4.1 billion are active, maximizing efficiency.

The context window supports 1 million tokens. Most layers use local attention, accessing only a short window of recent text, while only some layers use global attention to access the full context, reducing computational costs. Additionally, the positional encoding uses a method that was widely used in the past but is less preferred recently.

Image and audio inputs are integrated directly into the model without separate pre-trained encoders. It also provides a Thinking effort setting (0-1) to adjust the depth of reasoning before answering, allowing users to flexibly control the model's thought process according to their needs.