975 billion parameters! OpenAI, the former CTO’s star unicorn, released the first open weight model

Zhidongzhi Compiler | Eggplant Editor | Cheng Qian Zhidongzhi reported on July 16 that today, American AI unicorn Thinking Machines Lab founded by former CTO Mira Murati of OpenAI released its first multi-modal MoE open weight model Inkling. The core advantages of this model are that it can be transformed, deployed, and cost-controllable. Developers can fine-tune the model and adjust the inference intensity to control its effect, delay, and cost. Murati posted on the social platform X that this is our first model, Inkling. Training starts from scratch, the weights are open to the public, and fine-tuning can be carried out on the Tinker platform from now on. ▲Mira Murati issued a post announcing Inkling (Source: ▲Weng Li published an article introducing Inkling (Source: Thinking Machines Lab said that developers can adjust Inkling through Tinker, or they can first experience the chat and agent-style network search functions in the Inkling Playground, a test page of the Tinker console. At the same time, the complete weight of Inkling has been opened in Hugging Face, and developers can download the original checkpoint and NVFP4 checkpoint for deploying or adjusting the model. ▲Inkling open weights (Source: Hugging Face) According to the official blog, Inkling’s basic model draws on China’s DeepSee