Following the DeepSeek architecture, US large models start to copy China's "homework"
OpenAI's former CTO Murati has launched a "premium alternative" to a Chinese model.
The first model released by Thinking Machines, Inkling, mainly follows the DeepSeek-V3 Mixture-of-Experts architecture, and its post-training cold start also uses data generated by open models such as Kimi K2.5. However, in multiple benchmark evaluations, Inkling lags behind Kimi and GLM, while its invocation price is even higher.
For a star company that has raised $2 billion in financing, with a valuation of $12 billion, and whose team is packed with numerous former OpenAI employees, this result is somewhat below expectations.
But Murati probably never intended to compete for the top spot with Inkling from the very beginning. As U.S. enterprises face growing uncertainties when using Chinese models, a local open model with decent performance, customization support, and lower compliance risks has already carved out its own market. The "premium alternative" does not need to be overwhelmingly strong to find its own market.
Murati's First Model, Is This All?
On local time July 15, Thinking Machines Lab released Inkling, its first model since its founding.
Judging solely by its specifications, this model fully matches the grandeur that a star AI lab should possess.
Inkling adopts a Mixture-of-Experts architecture, with a total parameter count reaching 975 billion, activating 41 billion parameters for every Token processed; it uses 45 trillion Tokens for pre-training, with data covering text, images, audio, and video, and supports a maximum context length of 1 million Tokens. It can understand text, images, and audio, output results in text form, and also supports adjusting the "thinking intensity" to trade off between performance, speed, and cost.
Inkling has also made its model weights open, under a fairly permissive Apache 2.0 license. Developers can download the weights, deploy them on their own, or fine-tune the model through the Tinker platform under Thinking Machines.
The company has a clear positioning for it: instead of directly competing with ChatGPT for ordinary users, Inkling is more suitable to serve as a foundation for enterprises to develop AI applications.
However, the most notable part of Inkling is hidden in the second half of its technical report.
When introducing the model architecture, Thinking Machines explicitly wrote that the Mixture-of-Experts design of Inkling "mainly follows DeepSeek-V3". It also sets up a large number of expert modules, with only a small subset invoked for each Token, and adopts designs such as the auxiliary-loss-free load balancing from DeepSeek-V3.
The influence of Chinese models on Inkling extends to the training phase as well.
Thinking Machines stated that to kickstart post-training, the company first conducted a round of supervised fine-tuning using synthetic data generated by open-weight models, among which Moonshot AI's Kimi K2.5 was explicitly mentioned.
The entire technical path is now clear enough: an American lab founded by OpenAI's former CTO built its first model following DeepSeek's architecture, and completed the initial post-training phase with data generated by Kimi.
There is nothing inherently wrong with drawing on open models. DeepSeek and Kimi have made their weights public and allowed developers to use them, and scientific research and engineering development have always been built on the accumulation of publicly available achievements. The more awkward point is that after Inkling absorbed the technologies of Chinese models, its performance still did not surpass these "teachers".
Thinking Machines did not gloss over this point either. The company explicitly acknowledged in its release article: "Inkling is not the strongest-performing model available today, whether among open models or closed-source models."
According to the results published by Thinking Machines,
. 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 tool use capabilities, Inkling's score rose to 46%, still lower than Kimi K2.6's 54% and GLM 5.2's 54.7%.
. In SWE-Bench Pro, which assesses real-world software engineering capabilities, Inkling scored 54.3%, Kimi K2.6 58.6%, and GLM 5.2 reached 62.1%;
. The gap on Terminal Bench 2.1 is even more obvious: Inkling only got 63.8%, while Kimi K2.6 and GLM 5.2 hit 71.3% and 82.7% respectively.
Of course, Inkling also has its own strengths. It performs well in web application design, audio understanding, and safety evaluations, and can outperform Kimi and DeepSeek on some mathematical tasks.
However, in terms of comprehensive reasoning, programming, and agent capabilities, it is currently difficult to classify it into the first tier of open models.
Its performance did not reach the top tier, and its pricing also brought no surprises.
On the Tinker platform, the 64K version of Inkling charges $1.87 per million Prefill Tokens and $4.68 per million Sample Tokens. The former can be roughly understood as the input cost, while the latter is close to the generation cost, and this is already a "limited-time 50% off" price. The 256K version is even more expensive, reaching $3.74 and $9.36 respectively.
For comparison, Kimi K2.6's official API charges $0.95 per million input Tokens and $4 per million output Tokens, while GLM 5.2 charges $1.4 and $4.4 respectively.
Tinker also provides training and sampling services, and its billing method is not exactly the same as that of regular model APIs. But a rough comparison based on the above prices shows that Inkling's prefill price is nearly twice that of Kimi K2.6, and its generation price is also higher than that of Kimi K2.6 and GLM 5.2, with no price advantage visible for now.
As a result, Thinking Machines' first model presents a somewhat delicate situation: its architecture references DeepSeek, its post-training draws support from Kimi, its overall performance does not surpass top-tier Chinese open models, yet its invocation price is even higher.
Murati's Choice
Considering the background of Murati and Thinking Machines, this choice becomes even more thought-provoking.
If Inkling came from an ordinary AI startup, drawing on DeepSeek and using Kimi for training might not be so notable. But it comes from Thinking Machines — a company that can almost be called the "OpenAI Alumni Association".
When Thinking Machines officially debuted in February 2025, its team had only about 30 members, roughly two-thirds of whom came from OpenAI, with the rest mainly from Meta and Mistral.
The founding team once included OpenAI co-founder John Schulman, OpenAI's former VP of Research Barret Zoph, as well as researchers like Lilian Weng, Andrew Tulloch, and Luke Metz.
These individuals were involved in OpenAI's reinforcement learning, post-training, pre-training, safety, and reasoning research respectively. Reuters even directly described at the time that Murati poached at least 20 researchers from her former employer.
Standing among this group of OpenAI alumni, Murati is of course the most special one.
She joined OpenAI in 2018, participated in products such as DALL-E, Codex, ChatGPT, and Sora successively, and was promoted to CTO in 2022, responsible for coordinating research, model training, safety, and product implementation.
In November 2023, after the OpenAI board suddenly removed Sam Altman, Murati also briefly served as interim CEO. Her role was more akin to the overall leader of technology and products, with a complete understanding of OpenAI's model R&D process and productization path.
In other words, Thinking Machines never lacked OpenAI's experience from the very beginning of its founding.
Murati and her team know how a top closed-source lab develops models, and have personally experienced the entire process of ChatGPT evolving from a research achievement to a global product. When this group of people left OpenAI and built their own technical system from scratch, the outside world naturally expected them to come up with a path bearing the OpenAI imprint.
As it turns out, the clearest technical origin of Inkling points to China. This can easily be interpreted as a dramatic signal: the group of people who know OpenAI best left OpenAI, and ultimately chose the technical blueprint of Chinese models.
However, this conclusion needs to be analyzed in detail.
All of OpenAI's cutting-edge models in recent years have adopted a closed-source approach — the public cannot access the model weights, nor understand the complete architecture, training data, and post-training schemes. Even though Murati is familiar with OpenAI's internal technologies, she cannot directly bring her former employer's trade secrets into her new company. DeepSeek and Kimi, on the other hand, have made their weights or technical reports public, with their architectures, training methods, and toolchains legally available for research and reuse.
Inkling itself is also an open-weight model. For a company that wants to build an open model ecosystem, adopting mature solutions from open models such as DeepSeek and Kimi is a logical engineering choice.
To develop an open model today, completely avoiding the achievements that Chinese teams have already made public would actually lead to higher trial-and-error costs.
Therefore, the fact that Inkling draws on Chinese models does not directly prove that Murati believes DeepSeek's technology is comprehensively ahead of OpenAI. The two sides have different levels of openness, leading to completely different conditions for reference.
But this event still sends a clear signal: at least in the open-weight field, Chinese models have already become an important reference point for American AI teams.
Therefore, the choice to draw on Chinese models is not difficult to explain. The real question to answer is: Thinking Machines has absorbed ready-made architectures and training experience, why did the resulting Inkling end up with weaker performance and a higher price?
Murati has a top-tier team, sufficient capital, and NVIDIA's latest training systems. There is no reason for her to fail to see the performance and price gaps between Inkling and Chinese models.
Given that, this "premium alternative" is very likely a calculated choice.
"Decent Performance" Is Enough
Judging by the treatment Thinking Machines has received, Inkling was never supposed to be just a "good enough" model.
In July 2025, before the company launched any products, it completed a $2 billion seed round of financing, with a valuation of $12 billion. Its investors included a16z, NVIDIA, AMD, and Jane Street, among others. Four months later, Thinking Machines was reported to be in negotiations for a new round of financing, targeting a valuation of up to $50 billion. Although the transaction was never officially announced as completed, the outside world has already regarded it as a potential rival to OpenAI and Anthropic.
As a result, when Inkling debuted as a product with worse performance than Chinese models and a higher price, it inevitably seemed below expectations.
But from a business perspective, Murati probably never intended to use it to compete for first place in benchmark evaluations from the very beginning.
Thinking Machines emphasizes open weights, multi-modal capabilities, and customizability, hoping that enterprises will fine-tune it through Tinker to transform Inkling into specialized models for customer service, programming assistants, and industry agents. The market Murati is betting on is not so obsessed with taking first place in benchmark evaluations. Enterprises care more about whether the model can be privately deployed, whether they can have control over their data, and whether they can continue training the model based on their own business needs.
Previously, a large part of such demands were met by Chinese open models.
OpenAI and Anthropic in the U.S. have long adhered to a closed-source approach, and after Meta's Llama 4 underperformed expectations, it also gradually scaled back its open-source roadmap. Meanwhile, DeepSeek, Kimi, Qwen, and GLM have continued to make their weights public, with performance getting closer to that of American closed-source models, yet priced much lower.
American enterprises quickly voted with their feet.
According to data provided by OpenRouter to CNBC, since February 2026, Chinese models have accounted for more than 30% of the total Tokens invoked by U.S. enterprises via this platform every week, peaking at 46% at one point. The average proportion in the first half of 2025 was only 4.5%.
When developing Composer 2, Cursor tested multiple foundation models and ultimately chose Kimi K2.5, for the straightforward reason that it delivered the strongest performance in evaluations. Bridgewater Associates, on the other hand, fine-tuned Alibaba's Qwen through Tinker, and the resulting custom model outperformed some top-tier closed-source models at a lower cost.
Chinese open models have thus become a shortcut for American enterprises to reduce AI costs. However, this path is now under growing pressure.
At the same time, the U.S. regulatory atmosphere surrounding Chinese AI models is tightening. Relevant authorities have issued warnings about data security, model origins, and supply chain risks, and some enterprises using Chinese models have also been subjected to investigations. Even without unified restrictions in place yet, policy uncertainty is already enough to influence enterprise choices, especially for companies involved in government business and sensitive data.
This is exactly the market space for Inkling.
It is an open-weight model developed by an American company, under the Apache 2.0 license, supporting private deployment and customization. By choosing it, American enterprises do not need to worry about political controversies arising from using Chinese models, and can more easily pass compliance reviews from government clients and large corporations.
Inkling's real value comes from this kind of compliance certainty.
Murati has clearly spotted this emerging gap in the market.
Thinking Machines absorbed the public achievements of DeepSeek and Kimi, and then transformed them into a model provided by an American company. Inkling does not need to beat Chinese models in all benchmark evaluations — it only needs to become an acceptable option for American enterprises that dare not continue to use Chinese models.
This also explains why it can afford to be more expensive. Policy has narrowed Inkling's competitive landscape, making performance and price gaps far less fatal. For some American enterprises, a slightly weaker, slightly more expensive model with lower risks is already sufficient.
From this perspective, Inkling is even more proof that Murati understands how to do business.
However, the cost of this business model may be borne by American enterprises.
Companies outside the U.S. can still freely choose GLM, Kimi, and DeepSeek to develop products on stronger, cheaper foundation models; American enterprises, however, may be forced to adopt local alternatives with slightly weaker performance and higher prices. The capability gaps in foundation models will continue to be passed on to the fine-tuned products, eventually creating a structural competitive disadvantage.
Washington originally hoped to use restrictions to protect American AI companies, but the result may first create a market for "premium alternatives" like Inkling. For Murati, "decent performance" is already enough. For the American AI industry, this "good enough" standard means rising costs and declining competitiveness.
This article is from the WeChat official account "Letter AI", author: Xiao Jinya, published with authorization from 36Kr.