Ex-OpenAI CTO Mira Murati's Thinking Machine Releases First Model Inkling

Inkling adopts a Mixture of Experts (MoE) architecture, with total parameters as high as 975B (close to 1 trillion), but only 41B parameters are activated per inference, balancing scale and efficiency. The context window reaches 1M tokens, capable of handling extremely long texts. The training data scale is 45T tokens, covering multimodal content.

Additionally, Thinking Machine has also released a smaller preview model, Inkling Small, with 12B active parameters, targeting resource-constrained scenarios.

Inkling natively supports understanding and reasoning across three modalities: text, image, and audio, without the need for additional adaptation. The model weights are fully open source, allowing developers to freely download, fine-tune, or deploy.

Notably, its post-training data is not entirely self-developed but distilled from a series of open-source models, including Kimi 2.5. This approach reduces data acquisition costs but also raises discussions about originality and dependency.

Thinking Machine was founded by former OpenAI CTO Mira Murati in early 2025, and after a year and a half of development, it has finally launched its first product. Murati led key projects such as GPT-4 during her time at OpenAI, and the moves of her new company have been closely watched.

The release of Inkling marks the addition of a powerful player to the open-source AI model camp, especially in terms of MoE architecture and ultra-long context, competing with models like DeepSeek-V3 and Qwen2.5. However, Inkling's distilled data sources may spark discussions about model originality and intellectual property.