Mira Murati left OpenAI in September 2024. Less than two years later, her startup just shipped a model that makes most open-source alternatives look like science fair projects.
Thinking Machines Lab launched Inkling on July 15, a multimodal AI model packing 975 billion total parameters. It processes text, images, audio, and video. It supports a context window of up to 1 million tokens. And it’s available for free download on Hugging Face.
What Inkling actually is
Inkling uses a Mixture-of-Experts (MoE) architecture. Think of it like a massive office building where 975 billion employees work, but only 41 billion show up for any given task. The rest stay home. This design means you get the knowledge depth of a near-trillion parameter model without needing the compute budget of a small nation to run it.
The model was trained on 45 trillion tokens across multiple modalities. To put that in perspective, GPT-3 was trained on roughly 300 billion tokens. We’re talking about a training dataset roughly 150 times that size.
