Mira Murati, previously the Chief Technology Officer at OpenAI and now the CEO of Thinking Machines, recently spoke at the Wall Street Journal's WSJ Tech Live Conference held in Laguna Beach, California, on October 17, 2023. This marks a notable moment in her career following the founding of Thinking Machines Lab in February 2025, alongside several senior researchers from OpenAI. The company made headlines by raising an unprecedented two billion dollars in its seed round, achieving a remarkable twelve billion dollar valuation without releasing any products. Major investors include tech giants like Nvidia, AMD, Cisco, Andreessen Horowitz, and Jane Street, setting high expectations for the emerging firm.
This week, Thinking Machines unveiled its inaugural model, Inkling. It employs a mixture-of-experts transformer architecture featuring an impressive total of 975 billion parameters, with 41 billion being active per token. Inkling was trained on an extensive dataset of 45 trillion tokens, which encompasses text, images, audio, and video. It is designed to accommodate a context window of one million tokens. Despite these formidable numbers, the model's performance has not set it apart from its competitors. Thinking Machines readily acknowledges this, stating that Inkling is not currently the strongest model available, whether open or closed. Benchmark tests such as Humanity's Last Exam, Terminal Bench, and SWE-Bench Verified reveal that it lags behind leading Chinese open models, such as Zhipu's GLM 5.2 and Moonshot's Kimi K2.6.
What stands out is the fact that this twelve billion dollar American lab, staffed by the creators of ChatGPT, has introduced a model that does not outperform its free Chinese counterparts. A deeper investigation into this phenomenon reveals some intriguing insights.
For the past eighteen months, U.S. authorities have accused Chinese labs of basing their models on pilfered American technology. Following DeepSeek's impressive debut in early 2025, the rapid response from OpenAI and government officials was that DeepSeek had utilized distillation techniques on OpenAI models. Distillation is the process of training one's model using the outputs of a more advanced one. Subsequent Congressional inquiries have asserted that Chinese AI companies were engaging in coordinated distillation efforts aimed at U.S. frontier models, leading to accusations of theft.
With this context, let's delve into Inkling. Its architecture closely resembles that of DeepSeek-V3, a fact that both technical analysts and the Thinking Machines team itself have noted shortly after its release. The company admits that Inkling's post-training involved supervised fine-tuning with synthetic data derived from open-weight models, including the Chinese Kimi K2.5 model developed by Moonshot AI.
This raises an important point: an American lab has essentially embraced a Chinese model architecture and utilized a Chinese dataset for training its flagship product. While nobody is labeling this as theft—and rightly so, as the Chinese models in question are open-source and licensed for use—there exists a noticeable disparity in the narrative. When Chinese labs draw from American models, it’s considered criminal; however, when American labs learn from Chinese models, it is approached as engineering innovation. Such selective rhetoric poses challenges for coherent technology policy-making.
This brings us to a crucial question: why launch a model like Inkling, which isn't even the best in the market? The strategy appears calculated, benefiting from a protective market environment. Inkling is licensed under the open-weight Apache 2.0, predominantly marketed as a chatbot but also as a customizable platform through Tinker, the company’s fine-tuning tool. Thinking Machines seems to be banking on enterprises prioritizing adaptability over sheer capability.
However, could American companies not adapt superior Chinese models instead? They attempted to do so but found themselves gravitating towards Chinese open models, which are both high-quality and cost-effective. These Chinese models now represent about 45 percent of the enterprise tokens processed through OpenRouter. Companies like Coinbase have significantly cut down their AI expenditures by leveraging models like GLM and Kimi, while Cursor has developed its Composer model based on Kimi.
Despite this trend, the U.S. government is actively seeking to restrict access to these models. The State Department recently issued warnings to American businesses regarding the potential risks associated with Chinese models. Additionally, investigations have been initiated into firms like Airbnb and Cursor due to their use of Chinese AI tools. Proposed procurement bans are in the works, making it increasingly risky for any company involved with government contracts—or facing the prospect of congressional scrutiny—to engage with Chinese open models.
Inkling is positioned to fulfill the gap left by these restrictions. It serves as the compliant alternative, not the supreme one, but rather a sanctioned low-risk option.
Finally, it is essential to reflect on the background of Thinking Machines’ founder. Mira Murati’s tenure as OpenAI's CTO means she possesses unparalleled insight into the inner workings of OpenAI’s technology. Given her expertise and the significant resources available to her, her decision to adopt an architecture inspired by DeepSeek and a training strategy influenced by Kimi carries weight. This choice suggests a shifting landscape in the realm of open research, particularly with OpenAI experiencing a noticeable slowdown. The rollout of GPT-5.6 has been limited, while Kimi K2.6 has surpassed GPT-5.5 in certain benchmarks, marking a significant moment where an open-weight model has bested a leading proprietary model. The competitive landscape between American closed labs and their Chinese counterparts is tightening, with the distinctions between them narrowing on multiple fronts. When a former leader from OpenAI opts for a design derived from Chinese models instead of her previous employer’s methodologies, it compels the market to reevaluate assumptions about the strength of OpenAI’s position.
