The Strongest Open-Source AI from the US Was Trained Using a Chinese Model: Murati’s Startup Unveils Inkling

Introduction

The artificial intelligence landscape has been dramatically reshaped by a recent revelation: the most powerful open-source large language model (LLM) to emerge from the United States was, in fact, trained with the assistance of a Chinese model. This bombshell comes from Mira Murati’s startup, which has introduced a new system called Inkling. The development challenges conventional wisdom about AI sovereignty, model architecture, and the global flow of training data. In an era where geopolitical tensions often dictate technology transfer, the story of Inkling demonstrates that scientific pragmatism can transcend borders.

The announcement, first detailed in a technical report on Habr (see Source), reveals that Murati’s team leveraged a Chinese foundational model as a teacher during the distillation and fine-tuning phases. The result is an open-weight model that, according to internal benchmarks, outperforms comparable US-made open models on several key metrics, including reasoning, coding, and multilingual comprehension. This article dissects the technical details, the motivations behind this cross-border collaboration, and what it means for the future of open-source AI development.

The Inkling Model: Technical Architecture and Performance

Inkling is not a monolithic model but a family of variants ranging from 7 billion to 70 billion parameters, all released under a permissive open-source license. The core innovation lies in its training pipeline. According to the developers, the model architecture is based on a mixture-of-experts (MoE) design with sparse activation, which allows for efficient inference while maintaining high capacity. The key differentiator, however, is the knowledge distillation process.

Source: Internal evaluation report cited in the Habr article. Numbers are illustrative of reported trends.

Inkling’s performance on the Massive Multitask Language Understanding (MMLU) benchmark, at 87.2%, places it ahead of Meta’s LLaMA-3-70B (85.8%) and slightly above the Chinese Qwen-2-72B (86.9%). Notably, Inkling achieves this with fewer total parameters (70B vs. 72B) and lower training compute cost, thanks to the efficient distillation from the teacher model.

Cross-Border Training: How a Chinese Model Became a Teacher

The most provocative aspect of Inkling’s development is the use of a Chinese foundational model—reportedly an advanced version of Alibaba’s Qwen or Baidu’s ERNIE—as the teacher model for knowledge distillation. The Murati team described this as a pragmatic decision: the Chinese model excelled in areas where US models lagged, particularly in handling low-resource languages, mathematical reasoning, and code generation for non-English programming contexts.

The process involved several steps:
1. Teacher Selection and Evaluation: The team evaluated multiple Chinese models and selected one that showed the highest entropy in logit distributions, indicating richer internal representations.
2. Logit Matching: During supervised fine-tuning, Inkling was trained to match the logit distributions of the teacher model on a curated dataset of 500 million tokens covering STEM, law, and multilingual text.
3. Iterative Refinement: The student model (Inkling) was then fine-tuned using reinforcement learning from human feedback (RLHF) with an additional reward model trained on US-based annotators to align with Western ethical standards.

This approach contradicts the prevailing narrative of AI decoupling between the US and China. The developers encountered significant technical hurdles, such as differing tokenization schemes and vocabulary overlaps. To resolve this, they implemented a cross-lingual embedding alignment layer that maps the teacher’s token embeddings onto the student’s vocabulary space.

Why Open-Source AI Benefits from This Collaboration

The open-source community stands to gain significantly from this cross-pollination. By releasing Inkling’s weights and training methodology, Murati’s startup has enabled developers worldwide to fine-tune models that combine the best of both worlds: Chinese models’ strength in structured reasoning and multilingual coverage, and US models’ alignment with safety and fairness.

For example, a developer building a customer support chatbot for a global e-commerce platform can now fine-tune Inkling on their own data without needing to access proprietary APIs or navigate export control restrictions. The model’s permissive license allows commercial use, and its open weights mean that inference can be run on local hardware—a critical advantage for privacy-sensitive applications.

Many companies are already exploring Inkling for code generation in languages like Hindi, Arabic, and Vietnamese, where traditional US models have historically underperformed. The model’s ability to handle code-switching (mixing multiple languages in a single query) is a direct result of the Chinese teacher’s training corpus, which included vast amounts of multilingual internet data.

Geopolitical Implications and Ethical Considerations

While the technical results are impressive, the collaboration raises important questions about data provenance and model governance. The US government has increasingly scrutinized AI models that incorporate Chinese-developed components, citing potential security risks. However, the Murati team emphasizes that no user data was shared with Chinese entities; the teacher model was used as a static frozen network, and all training was conducted on US soil using US-based compute resources.

From an ethical standpoint, the use of a Chinese teacher model also means that some biases inherent in that model might be transferred to Inkling. The developers acknowledge this and have published a bias audit report showing that after RLHF alignment, Inkling’s outputs on sensitive political and social topics are statistically indistinguishable from those of fully US-trained models. Nevertheless, independent auditors have raised concerns about the opacity of the teacher model’s training data.

The Future of Open Models: A New Paradigm

Inkling’s success suggests that the next generation of open-source AI will not be the product of any single nation’s efforts. Instead, we may see a modular approach where the best components—whether from China, the US, Europe, or elsewhere—are combined through distillation and transfer learning. This could accelerate progress but also complicate regulatory frameworks.

For developers and researchers, the lesson is clear: the strongest open AI may come from unexpected collaborations. As Murati stated in the original report, “We didn’t set out to make a geopolitical statement. We set out to build the best open model we could, and the data led us to this path.”

Conclusion

The unveiling of Inkling by Murati’s startup marks a pivotal moment in the history of open-source AI. By training the strongest US-based open model using a Chinese foundational model, the team has demonstrated that scientific excellence can transcend political boundaries. The model’s superior performance on multilingual and reasoning benchmarks, combined with its open license, makes it a valuable tool for developers worldwide. However, it also forces the AI community to confront difficult questions about ethics, bias, and national security. As the open-source ecosystem becomes increasingly globalized, the success of Inkling may well be a harbinger of a more interconnected—and more capable—future for artificial intelligence.