Inkling, Built on Chinese Blueprint, Exposes US AI Policy Asymmetry

Thinking Machines Lab, founded by former OpenAI Chief Technology Officer Mira Murati, released its first model this week. The company raised $2 billion at a $12 billion valuation before shipping a product. Nvidia, AMD, Cisco, Andreessen Horowitz and Jane Street were among the backers.

The 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 impressive numbers. Yet the performance lags leading Chinese open models.

Thinking Machines itself concedes this. On benchmarks like Humanity's Last Exam, Terminal Bench and SWE-Bench Verified, Inkling trails Zhipu's GLM 5.2 and Moonshot's Kimi K2.6, the company said.

What is revealing is how the model was built. Inkling's mixture-of-experts architecture largely follows DeepSeek-V3. Technical analysts noted this within hours of the release. Thinking Machines also acknowledged that Inkling's post-training used synthetic data generated by open-weight models, including Kimi K2.5 from Moonshot AI. In other words, an American lab adopted a Chinese architecture and distilled a Chinese model to train its flagship release.

"When Chinese labs learn from American models it is larceny. When American labs learn from Chinese models it is engineering."

Washington has spent eighteen months accusing Chinese labs of stealing American intellectual property. Congressional investigators claim Chinese AI firms ran coordinated distillation campaigns against US frontier models. The word used is theft. The same label has not been applied to Thinking Machines. The Chinese models are open weight and permissively licensed. Learning from published work is how science advances. The asymmetry in rhetoric is now visible.

The obvious question is why release Inkling at all. The answer is found in a strategy that benefits from protectionism. Inkling is open weight under Apache 2.0. It is being pushed as a base for customization through Tinker, the company's fine-tuning platform. Thinking Machines bets that enterprises care less about the smartest general model than about a model they can make their own.

American companies already flocked to Chinese open models. They are excellent and nearly free. Chinese models account for roughly 45 percent of enterprise tokens routed through OpenRouter, the analysis notes. Coinbase cut its AI bill nearly in half by moving its agents to GLM and Kimi. Cursor built its Composer model on Kimi.

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. The risk overhang is massive.

Inkling exists to fill that vacuum. It is the compliant base model. Not the best base model. Just the permitted, low-risk one.

The center of gravity in open AI has shifted to China. 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, the market should update its beliefs about how large OpenAI's remaining moat really is.

For US tech companies, the policy response creates a structural disadvantage. A fine-tuned model inherits the ceiling of its base. If a firm in Singapore or Berlin 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. Because policy walls them off from the best available inputs while the rest of the world builds freely.

Nvidia, which supplies the GB300 systems used to train Inkling, is central to AI compute. Its NVDA stock page carries an Alpha Score of 74. AMD stock page, also a Thinking Machines backer, has a score of 58. Both face a market where the best open models come from China, potentially altering demand dynamics.

The House Science Committee has launched probes into companies like Airbnb and Cursor over their use of Chinese AI. A formal procurement ban has not been drafted. The committee plans to mark up the bills in the coming weeks.