Former CTO of OpenAI Launches New Company, Open Sources First Model: Inkling, Architecture Inspired by DeepSeek-V3

According to Dynamic Beating monitoring, the Thinking Machines Lab founded by former OpenAI CTO Mira Murati has released its first universal model, called Inkling. The model has a total of 975 billion parameters, with each activation requiring 41 billion parameters. Inkling supports text, image, and audio inputs with a maximum context length of 1 million tokens.

Inkling was trained from scratch, with the MoE architecture primarily based on DeepSeek-V3. During the later stages of training, data generated by open models such as Kimi K2.5 was also utilized. The complete weights have been uploaded to Hugging Face and are licensed under Apache 2.0.

Thinking Machines explicitly acknowledges that Inkling is not currently the strongest model in terms of performance. Its main focus is on customization and efficiency. When achieving similar performance on Terminal Bench 2.1, the number of tokens used is approximately one-third that of the NVIDIA Nemotron 3 Ultra.