OpenAI's former Chief Technology Officer Murati's first work after leaving her former employer has unexpectedly reversed the roles of China and the US in large model development. On July 15, she founded Thinking Machines Lab and released its first model, Inkling. The hybrid expert architecture mainly follows DeepSeek-V3, and for post-training cold start, it also used synthetic data generated by open models such as Moonshot's Kimi K2.5. The results are quite subtle: In multiple benchmark evaluations, Inkling lags behind Kimi and GLM, and its calling price is even higher. A star company that raised $2 billion and was valued at $12 billion, its debut was somewhat below expectations, but Murati may have never intended to compete for the top spot from the beginning.

Looking at the hardware configuration alone, Inkling indeed lives up to the image of a star lab. It uses a hybrid expert architecture with a total parameter count of 975 billion, activating 41 billion parameters per token processed; pre-training used 45 trillion tokens, covering text, images, audio, and video, supporting up to 1 million tokens of context; it can understand text, images, and audio, outputting in text form, and supports adjusting thinking intensity to balance performance, speed, and cost. The model also opens its weights, using a relatively liberal Apache 2.0 license, allowing developers to download weights for self-deployment or fine-tune through Thinking Machines' Tinker platform. The company's positioning is clear: rather than directly competing with ChatGPT for ordinary users, Inkling is more suitable as a foundation for enterprises developing AI applications.

What really caught attention is hidden in the latter part of the technical report. Thinking Machines explicitly stated in the architecture introduction that Inkling's hybrid expert design mainly followed DeepSeek-V3, with a large number of expert modules, where each token only activates a small portion, and it also adopted DeepSeek-V3's no-assistant loss load balancing design. The influence of Chinese models extended into the training phase: to initiate post-training, the company first performed a round of supervised fine-tuning using synthetic data generated by open-weight models, explicitly mentioning Kimi K2.5.

A complete roadmap emerges— an American laboratory founded by OpenAI's former CTO, its first model follows DeepSeek architecture, and then starts post-training with Kimi's data. Borrowing from open models itself is not problematic, as DeepSeek and Kimi publicly release weights and allow usage, which is built upon public achievements. More awkwardly, after absorbing the technology of Chinese models, Inkling still did not surpass these teachers. Thinking Machines also candidly admits: Inkling is not currently the strongest model, whether open or closed-source.

When the numbers are laid out clearly, it becomes even clearer. In the final human exam pure text evaluation, Inkling scored 29.7%, while Kimi K2.6 and GLM 5.2 reached 35.9% and 40.1%, respectively; after adding tools, Inkling rose to 46%, still lower than Kimi K2.6's 54% and GLM 5.2's 54.7%; in SWE-Bench Pro, which tests real software engineering capabilities, Inkling scored 54.3%, while Kimi K2.6 was 58.6%, and GLM 5.2 reached 62.1%; the gap was most glaring in Terminal Bench 2.1, where Inkling scored 63.8%, while Kimi K2.6 and GLM 5.2 were 71.3% and 82.7%, respectively.

Of course, Inkling has strengths in web application design, audio understanding, and security evaluation, and it can even exceed Kimi and DeepSeek in some math tasks. However, in terms of comprehensive reasoning, programming, and agent capabilities, it hasn't yet made it into the top tier of open models. Its pricing also didn't bring surprises: on the Tinker platform, the 64K version charges $1.87 per 1 million Prefill Tokens and $4.68 per 1 million Sample Tokens, and this is already a five-day discounted price. The 256K version rises to $3.74 and $9.36, respectively; compared to Kimi K2.6's official API input of $0.95 and output of $4, and GLM 5.2's $1.4 and $4.4, Inkling's prefill price is nearly twice that of Kimi K2.6, and its generation price is also higher, showing no obvious price advantage. Thus, the situation becomes a bit delicate: the architecture references DeepSeek, the post-training relies on Kimi, the overall performance doesn't exceed China's leading open models, and the calling price is even higher.

The reason this is intriguing lies in the unique background of Thinking Machines. The company is almost an OpenAI alumni association— when it was introduced in February 2025, the team had about 30 people, two-thirds came from OpenAI, and the rest mainly from Meta and Mistral. The founding team once included John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz, and Reuters reported that Murati had recruited at least 20 researchers from her former employer.

Murati joined OpenAI in 2018, participated in DALL-E, Codex, ChatGPT, and Sora, was promoted to CTO in 2022, and briefly served as interim CEO after the board dismissed Altman in November 2023, gaining full insight into OpenAI's model development and productization path. Such a group of people who know OpenAI best left to build a technical system from scratch, and the outside world expected a route marked by OpenAI's imprint. However, the most obvious technical source turned out to point to Chinese models, which could easily be interpreted as a dramatic signal: the people who knew OpenAI best ultimately chose the technical blueprint of Chinese models.

But this conclusion needs to be analyzed carefully. OpenAI's recent cutting-edge models are all closed-source, so the outside world cannot obtain weights or understand the complete architecture and training plan. Even if Murati was familiar with internal technologies, she couldn't transfer OpenAI's commercial secrets into the new company; however, DeepSeek and Kimi have publicly released weights or technical reports, and their architectures, training methods, and toolchains can be legally researched and reused.

Inkling itself is also an open-weight model, and for companies wanting to build an open ecosystem, adopting mature solutions from open models like DeepSeek and Kimi is a logical engineering choice. Today, developing an open model while completely avoiding the publicly available achievements of Chinese teams would require a higher cost of trial and error. Therefore, drawing inspiration from Chinese models cannot prove that Murati believes DeepSeek is superior to OpenAI. The differences in openness and the conditions for reference are completely different between the two, but it sends a clear signal: at least in the field of open weights, Chinese models have become an important reference for American AI teams.

The real question is why Thinking Machines, having absorbed existing architectures and training experiences, produced an Inkling with weaker performance and higher prices. Murati has a top team, sufficient funds, and the latest NVIDIA training system. There is no reason for her to overlook these gaps, so the high-cost alternative might be a calculated choice.

Considering the treatment Thinking Machines received, Inkling shouldn't just be a decent model. In July 2025, the company completed a $2 billion seed round before launching any product, with a valuation of $12 billion, and investors include a16z, NVIDIA, AMD, and Sequoia. Four months later, the next round of negotiations aimed for a valuation of up to $500 billion, although not officially announced, the company was already seen as a potential competitor to OpenAI and Anthropic. When it launched a product that performed worse than Chinese models and was even more expensive, it inevitably fell short of expectations. But from a business perspective, Murati may have never intended to win the benchmarks from the beginning.

The company emphasizes open weights, multimodality, and customizability, hoping that enterprises will use Tinker to fine-tune Inkling and turn it into specialized models such as customer service, coding assistants, and industry agents. Murati's bet is on a market that does not prioritize benchmark rankings. Enterprises care more about whether the model can be privately deployed, whether they can control the data, and whether they can continue training according to their own business. A large portion of this demand has been met by Chinese open models: OpenAI and Anthropic have long been closed-source, and after Meta's Llama4 did not meet expectations, it gradually reduced its open-source approach. Meanwhile, DeepSeek, Kimi, Qwen, and GLM continuously released weights, with performance increasingly close to American closed-source models and much lower prices.

US enterprises have quickly voted with their feet— according to data provided by OpenRouter to CNBC, since February 2026, the weekly share of Chinese models in tokens called by US enterprises through the platform has exceeded 30% every week, reaching as high as 46%, while the average share in the first half of 2025 was only 4.5%. Cursor tested multiple base models when developing Composer2 and finally chose Kimi K2.5, which had the strongest evaluation. Bridgewater Fund also fine-tuned Alibaba's Qwen through Tinker, achieving better results than some top closed-source models. Thus, Chinese open models have become a cost-saving shortcut for US enterprises, but this path is now under pressure.

At the same time, the regulatory atmosphere in the US toward Chinese AI models is tightening. Relevant departments have issued warnings about data security, model sources, and supply chain risks, and some companies using Chinese models have faced investigations. Even without unified restrictions, policy uncertainty has been enough to affect enterprise choices, especially for companies involving government business and sensitive data.

This is exactly the market space for Inkling: it is an open-weight model developed by an American company, using the Apache 2.0 license, supporting private deployment and customization. US enterprises choose it, without worrying about political controversies caused by using Chinese models, and it is easier to pass compliance reviews from government clients and large companies. The real value of Inkling comes from this compliance certainty— Murati transformed the public achievements of DeepSeek and Kimi into a model provided by an American company. It doesn't need to defeat Chinese models in benchmarks, but only needs to be an acceptable choice for US enterprises that dare not continue using Chinese models. This also explains why it dares to set a higher price: the policy has narrowed the competitive scope, and the gap in performance and price is less fatal. For some US enterprises, a slightly weaker and more expensive model is already sufficient.