Thinking Machines Lab released its first model this week. The company was founded in February 2025 by Mira Murati, the former Chief Technology Officer of OpenAI, along with a cohort of senior OpenAI researchers. It raised a record two billion dollar seed round at a twelve billion dollar valuation before shipping a single product. Nvidia, AMD, Cisco, Andreessen Horowitz and Jane Street backed it. Expectations are enormous.
The newly released model is called Inkling. It is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per token. It was pretrained on 45 trillion tokens spanning text, images, audio and video. It supports a one million token context window. These are very impressive numbers, yet the performance in comparison with other models is less impressive. To their credit, Thinking Machines concede this themselves. In the company’s own words, Inkling is not the strongest model available today, open or closed. On benchmarks like Humanity’s Last Exam, Terminal Bench and SWE-Bench Verified, it trails leading Chinese open models including Zhipu’s GLM 5.2 and Moonshot’s Kimi K2.6.
It is quite remarkable that a twelve billion dollar American lab staffed by the people who built ChatGPT have released a model that underperforms free Chinese alternatives. It is worth diving into, and when we do, the details end up being quite revealing.
Distillation Wars
For eighteen months Washington has accused Chinese labs of building their models on stolen American intellect. When DeepSeek stunned the AI community in early 2025, the immediate response from OpenAI and from US officials was that DeepSeek had distilled OpenAI’s models. Distillation means using a stronger model’s outputs to train your own. Congressional investigators have since claimed that Chinese AI firms ran coordinated distillation campaigns against US frontier models. The word used in Washington is theft.
With that background in mind, let’s take a look at Inkling. Its mixture-of-experts design largely follows DeepSeek-V3. That is not my characterization. It is plainly visible in the published architecture and has been noted by technical analysts within hours of release. And by Thinking Machines’ own admission, Inkling’s post-training was bootstrapped with supervised fine-tuning on synthetic data generated by open-weight models including Kimi K2.5, a Chinese model from Moonshot AI.
In other words, an American lab adopted a Chinese architecture and distilled a Chinese model to train its flagship release. Nobody is calling this theft. And frankly, nobody should. The Chinese models in question are open weight and permissively licensed. Learning from published work is how science advances. But the asymmetry in rhetoric is now impossible to ignore. When Chinese labs learn from American models it is larceny. When American labs learn from Chinese models it is engineering. You cannot run a serious technology policy on that kind of selective outrage.
So Why Launch A Model That Isn’t The Best?
The obvious question is why release Inkling at all. The answer is found in a shrewd strategy that benefits from protectionism. Inkling is open weight under Apache 2.0. It is being pushed primarily as a chatbot but as a base for customization through Tinker, the company’s fine-tuning platform. Thinking Machines is betting that enterprises care less about the smartest general model than about a model they can make their own.
But can American companies not do that with better Chinese models? They tried to. American companies flocked to Chinese open models because they are excellent and nearly free. Chinese models came to account for roughly 45 percent of enterprise tokens routed through OpenRouter. Coinbase cut its AI bill nearly in half by moving its agents to GLM and Kimi. Cursor built its Composer model on Kimi.
But Washington is now moving to shut that door. The State Department warned American companies this month about the risks of Chinese models. House committees launched probes into Airbnb and Cursor over their use of Chinese AI. Procurement bans are being drafted. For any company that touches government work, and increasingly for any company that fears congressional subpoenas, Chinese open models are becoming untouchable regardless of what the final rules say. The risk overhang is massive, and that is all that’s needed to push companies in other directions.
Inkling exists to fill that vacuum. It is the compliant base model. Not the best base model. Just the permitted, low-risk one.
What Inkling’s Architecture Tells Us About OpenAI
Let’s revisit for a moment who the founder of Thinking Machines is. Mira Murati was OpenAI’s CTO. She knows exactly what is inside OpenAI’s stack. She served as interim CEO during the November 2023 board crisis. Few people on earth have better visibility into what OpenAI has built and what it can build. Given a blank slate, two billion dollars and Nvidia’s newest GB300 systems, she chose an architecture that follows DeepSeek and a training bootstrap that leans on Kimi.
That is a not insignificant data point about where the technical frontier of open research actually sits. It arrives alongside a measurable deceleration at OpenAI itself. GPT-5.6’s rollout was limited. Kimi K2.6 edged past GPT-5.5 on SWE-Bench Pro, the first time an open-weight model surpassed a leading proprietary model on that benchmark. The gap between American closed labs and Chinese open labs has compressed to months on many tasks and has inverted on some. When the person who ran OpenAI’s technology organization builds from Chinese blueprints rather than from anything resembling her former employer’s approach, the market should – like any good Bayesian thinker – update its beliefs about how large OpenAI’s remaining moat really is.
AI Protectionism Can Hurt America
Let’s summarize our findings thus far. Inkling underperforms the best Chinese open models. American companies are being pushed by warnings, probes and coming regulation toward American bases like Inkling. Companies outside the United States face no such pressure. They will fine-tune GLM, Kimi and DeepSeek, the strongest open foundations on earth, while their American competitors build on weaker permitted alternatives.
Think about what that means. A fine-tuned model inherits the ceiling of its base. If a firm in Singapore, Istanbul or Barcelona starts from a stronger foundation than a firm in Austin or New York, the derivative products will reflect that difference. For the first time in the history of computing, American companies may operate at a structural disadvantage in a foundational technology, not because they lack talent or capital but because policy walls them off from the best available inputs while the rest of the world builds freely.
I have argued for years that the US-China AI competition will be decided by ecosystems, not by individual pieces, and that placing curbs on US companies making full use of the global market and all it produces is counterproductive. Open weights are becoming the substrate of this ecosystem. Chinese players seem to understand this very well and have flooded the world with capable free models. America’s frontier labs continue to keep their best work closed, and now the policy response to Chinese openness is to restrict American builders rather than out-compete on openness. Inkling is a genuine and welcome step toward an American open ecosystem. But it is also that it benefits from Chinese research and, for now, seems to need protectionism in order to justify its value. Meanwhile, Chinese open-weight models are closing in on frontier US capabilities.
The lesson of Inkling is not that Thinking Machines failed in any way, or is doing anything wrong. It is that the center of gravity in open, free-as-in-freedom AI has shifted to China. So much so that even OpenAI’s own alumni are building on Chinese foundations. Washington’s instinct to wall off American industry may end up handicapping the very companies it intends to protect. Strength comes from building better, faster, and yes, cheaper. Not from forcing your own team to play with weaker equipment.
