OpenAI's first open-source model, Inkling, based on DeepSeek-V3

The first general model, Inkling, was launched by Thinking Machines Lab, former CTO Mira Murati of OpenAI. The total parameter of the model is 97.5 billion, with 41 billion activated each time, supporting text, image and audio, with a maximum context of 1 million tokens. Inkling was trained from zero, but its MOE structure was mainly based on DeepSeek-V3, and later training used data generated from open models such as Kimi K2.5. Full weight has been put on Hugging Face, with Apache 2.0 permission. Thinking Machines explicitly acknowledges that Inkling is not currently the strongest model, but is primarily customized and efficient. When Terminal Bench 2.1 reached the same level of achievement, the token used was about one third of Nemotron 3 Ultra in British Weidar。