The American big model copied the work of DeepSeek: the performance was unbeatable and the price was more expensive, but it accurately blocked the compliance needs of American companies. The first answer sheet handed over by OpenAI’s former chief technology officer Mulati after leaving his old employer unexpectedly reversed the role relationship between Chinese and American big models.

American big models copied the work of DeepSeek: the performance was unbeatable and the price was more expensive, but it accurately blocked the compliance needs of American companies. The first answer sheet handed over by OpenAI’s former chief technology officer Mulati after leaving his old employer unexpectedly reversed the role relationship between Chinese and American big models. On July 15, the Thinking Machines Lab founded by her released its first model Inkling. The hybrid expert architecture mainly uses DeepSeek-V3, and the post-training cold start also borrows synthetic data generated by open models such as Dark Side of the Moon Kimi K2. 5. The results are quite subtle: Inkling lags behind Kimi and GLM in multiple benchmarks, and the call price is actually higher. A star company that raised $2 billion and was valued at $12 billion debuted somewhat below expectations, but Mulati may not have intended it to compete for first place from the start. Looking at the hardware configuration alone, Inkling can indeed support the ranking of a star laboratory. It adopts a hybrid expert architecture with a total parameter volume of 975 billion, and 41 billion parameters are activated for each Token processed; 45 trillion Tokens are used for pre-training, and the data covers text, images, audio and video, and supports up to 1 million Token contexts; it can understand text, pictures and audio, and output it in text form. It also supports the trade-off between performance, speed and cost by adjusting the intensity of thinking. The model also opens up the weights and adopts a fairly loose Apache 2. 0 license. Developers can download the weights and deploy them themselves, or fine-tune them through the Tinker platform of Thinking Machines. The company's positioning is clear: instead of directly competing with ChatGPT for ordinary users, Inkling is more suitable as a base for enterprises to develop AI applications. What is really eye-catching is hidden in the second half of the technical report. Thinking Machines wrote directly when introducing the architecture that Inkling's hybrid expert design mainly follows DeepSeek-V3. It also sets up a large number of expert modules, each Token only calls a small part of them, and adopts the unassisted loss load balancing design of DeepSeek-V3. The influence of the Chinese model also extends to the training stage: to launch post-training, the company first did a round of supervised fine-tuning using synthetic data generated by the open-weight model, which explicitly mentioned Kimi K2.5. A complete route has emerged - an American laboratory founded by the former CTO of OpenAI. The first model follows the DeepSeek architecture and starts training with the help of Kimi's data. There is nothing wrong with learning from the open model itself. DeepS