Jensen Huang's First Tweet Supports Open-Source AI Amid 'Kimi Panic'

On July 16, Moonshot AI released and is about to open-source the large model Kimi K3, with a total of 2.8T parameters, the largest model accessible in the open-source community, making breakthrough innovations in underlying architecture, with programming capabilities close to Fable 5. After release, coding packages and web interfaces were sold out or rate-limited, Hugging Face even set up a countdown page for K3, and overseas tech communities circulated 'Deployment Preparation Guides.'

The launch of Kimi K3 triggered a sense of threat from OpenAI and Anthropic, who accused Chinese companies of using model distillation to steal intellectual property and lobbied U.S. officials. The U.S. Treasury and other departments have begun considering measures such as banning some Chinese open-source AI models. U.S. tech media Axios called this chain reaction 'Kimi Panic.'

Meanwhile, nearly 200 companies and investment institutions formed the 'Small Tech Alliance,' warning the government that if cheap open-source models are banned, startups will be forced to pay high API fees to OpenAI and Anthropic, directly destroying many capital-limited U.S. startups.

Model distillation is an AI training technique where a pre-trained large model acts as a teacher, and a smaller model learns from the teacher model's outputs, thereby gaining capabilities close to the teacher model at lower cost. OpenAI officially launched a model distillation tool in 2024. Professor Li Meng of Nanjing University points out that the essence of distillation is capability transfer between models, evolving from the initial 'model compression' to today's 'capability replication.'

Black-box distillation in the era of large models typically involves submitting a large number of questions via API, collecting answers, code, and other data, then performing supervised fine-tuning on the student model. The student model does not copy the teacher model's weights, source code, or internal structure, but learns its output behavior. A Chinese scholar at a Canadian university noted that foreign media reports mentioning Kimi K3 'reconstructing model code and thought processes' are not rigorous, because API outputs cannot reverse-engineer the original model's source code or parameters.

The boundary between distillation and intellectual property infringement depends on multiple factors: whether the license of the teacher model's outputs allows use for training other models; whether obtaining outputs violates terms of service, geographical restrictions, or uses fake accounts, etc. The U.S. Copyright Office's analysis of AI training did not conclude that 'training is always legal' or 'always infringing,' requiring case-by-case judgment.

U.S. Treasury Secretary Bessent claimed to have found watermark evidence of U.S. large models in Chinese models, but technical details were not disclosed. The Chinese scholar in Canada said that foreign media reports did not disclose the specific type of watermark, detection protocol, sample size, threshold, or false positive rate, making independent evaluation of evidence strength impossible. Watermark techniques include various types, such as output text statistical watermarks, trigger-based model watermarks, and behavioral fingerprints.

Some model vendors embed encrypted 'thought signatures' in streaming outputs, containing the complete original chain of thought, allowing comparison of reasoning paths to determine whether the training data source with watermarks was accessed. However, such results can only prove that the training process may have accessed that data source, not independently prove infringement. Serious technical evidence should disclose more details.

Watermarks can detect the existence of distillation behavior, but it is difficult to prove that distillation is the main reason Kimi K3 achieved its current level. A model can simultaneously use multiple training methods, with distillation being just one. Currently, there is no unified technical standard for whether distillation constitutes intellectual property theft.

Public technical work by Chinese teams in recent years shows that China has achieved relatively high autonomy in model algorithms, training engineering, inference optimization, and open-source ecosystems. For example, DeepSeek-V3 disclosed designs like MLA and DeepSeekMoE; DeepSeek-R1 demonstrated a route of large-scale reinforcement learning from a base model; Kimi K1.5 disclosed methods like long-context reinforcement learning. These are clearly not obtainable by simply 'calling foreign APIs.'

If future restrictions only target U.S. frontier model APIs, the most direct losses for Chinese companies would be high-quality synthetic data, teacher models, and automatic evaluators. Post-training stages like math, code, and agent tasks would slow down, with the biggest impact on small and medium teams lacking their own teacher models. However, already trained and downloaded weights can still run. If restrictions further cover chips, cloud computing power, model distribution, etc., the impact would be greater.

The most important technical preparation is to establish an independent R&D chain, including proprietary flagship large models, verifiable reward training, independent evaluation systems, and advancing low-precision training, MoE, communication optimization, and domestic accelerator adaptation. The success of Kimi K3 proves that under constrained computing power, China can still produce models close to world-leading levels.