OpenAI's ex-CTO just open-sourced her first model, Inkling — and it borrowed from DeepSeek-V3

Mira Murati, former CTO of OpenAI, and her new company Thinking Machines Lab have released their first general-purpose model, Inkling. The model has 975 billion total parameters with 41 billion activated per inference, handles text, images, and audio, and supports a context window of up to 1 million tokens.

Inkling was trained from scratch, but its Mixture-of-Experts architecture primarily draws inspiration from DeepSeek-V3. During early post-training, the team also used synthetic data generated by open models like Kimi K2.5. The full model weights are available on Hugging Face under an Apache 2.0 license.

Thinking Machines openly acknowledges that Inkling isn't the most powerful model out there. Its selling points are customizability and efficiency. On the Terminal Bench 2.1 benchmark, Inkling achieves the same score while using roughly one-third the tokens required by Nvidia's Nemotron 3 Ultra.