Anthropic Stands Alone Against Open Models

The success of Chinese open-weight AI models has reignited a debate in Silicon Valley over whether America’s AI future should remain closed or embrace openness. Policymakers are grappling with this question through complex deliberations spanning national security, intellectual property, Chinese AI models, export controls, distillation, and the US’ strategy for maintaining its technological lead.

Recently, 50 companies and organisations, including NVIDIA, Microsoft, Meta, IBM, Dell Technologies, Mozilla, Palantir, Mistral, and Perplexity, signed a public letter urging US lawmakers not to impose restrictions on open-weight AI models. Noticeably absent were Anthropic and Amazon, whose most capable models remain proprietary.

Separately, NVIDIA, along with more than 40 technology companies, has launched the Open Secure AI Alliance, an industry initiative focused on developing open technologies for AI safety and cybersecurity.

It argues that open models, open agent harnesses (the underlying software infrastructure that turns LLMs into agents), and security tools should be treated as defensive infrastructure rather than security risks. 

“The world needs both closed and open models,” the alliance said in its announcement. “For cybersecurity, open models and open harnesses are essential because they democratise defensive capabilities, increase transparency for defenders, enable cyber defence while protecting data, and complement frontier closed models with customisable, localised controls.”

Much of this debate has been driven by China’s rapid progress. The country has recently accelerated its push into open-weight AI with models such as Kimi K3 and GLM 5.2, whose performance is increasingly approaching that of US frontier models such as Fable 5 and GPT-5.6.

The alliance also cited the recent Hugging Face security incident as an example of why open AI infrastructure is needed. According to the alliance, Hugging Face relied on the open-weight GLM 5.2 model running on its own infrastructure to analyse more than 17,000 actions during the incident after closed AI services were unable to support forensic analysis.

That said, the rise of Chinese models has not gone down well in Washington, where officials fear China could erode America’s lead in AI. Those concerns came into sharp focus last week when the US accused Chinese AI startup Moonshot AI of distilling Anthropic’s Fable model to train Kimi K3.

White House Office of Science and Technology Policy Director Michael Kratsios alleged that Moonshot had developed a sophisticated internal platform to carry out large-scale distillation of US frontier models, allowing it to switch between multiple methods of access to avoid detection.

Distillation in LLMs is a training process where a large, powerful model teaches a smaller model to copy its behaviour.

Jensen Steps Up

No executive has embraced the open-model argument more aggressively than NVIDIA CEO Jensen Huang. So much so that his first post on X was a letter explaining why open models matter. Other tech leaders, including Microsoft CEO Satya Nadella and Google CEO Sundar Pichai, also publicly voiced support for open models.

Huang believes open models strengthen, rather than weaken, the AI ecosystem. In an interview with Axios, he argued that open models increase security because researchers can continuously examine them for vulnerabilities.

“If everything just becomes one single model, one single point of attack, one single source of failure, I think the world is much, much more vulnerable,” Huang said.

He also challenged arguments that Chinese open models should be prohibited in the United States.

“These Chinese models are excellent,” Huang said. “Open-source models that are excellent should be used.” He added that there was ‘zero possibility’ China would drive American AI companies out of business simply by releasing competitive open models.

Huang has also argued that restricting companies like Anthropic from deploying powerful models does little to improve security because capable open models already exist globally.

He further said that US companies should use Chinese models without hesitation. Companies could download the models, fine-tune them, and build guardrails as they see fit.

“These models don’t just operate. They operate within what is called a harness. And that harness sits within what is essentially called a sandbox. These sandboxes are secure,” he said.

Google DeepMind CEO Demis Hassabis also supported the open-source moment. “A strong and secure open ecosystem is important for the world to benefit from AI. We’ve always supported and contributed heavily to open source and science from Jax to Transformers to AlphaFold to Gemma open models, which have now been downloaded 300M+ times,” he wrote on X.

While OpenAI has signed the letter urging lawmakers not to restrict open-weight models, OpenAI Head of Strategic Futures Dean Ball has offered a more critical view of Chinese open-weight models. While praising Moonshot AI's Kimi K3 as ‘a very good model’ whose performance ‘can't be explained away by distillation’, Ball warned that a world dominated by open-weight AI could lead to AI becoming a state-provided public good rather than a commercial product.

To Distil or Not to Distil

Anthropic remains one of the few major AI companies resisting the push towards open-weight models. The company argues that making frontier models widely available would make it easier for rivals to distil their capabilities and erode the competitive advantage of US AI labs.

It has accused Chinese labs of using unauthorised distillation to accelerate their AI development. The company argues that while distillation is a legitimate technique when used on a developer's own models, it becomes intellectual property theft when competitors extract capabilities from proprietary systems without permission.

In February, Anthropic disclosed what it described as ‘industrial-scale’ distillation campaigns involving DeepSeek, Moonshot AI, and MiniMax. The company said the three organisations collectively generated more than 16 million interactions with Claude through about 24,000 fraudulent accounts, allegedly using proxy networks and fake identities to bypass access restrictions and evade detection.

Most recently, Claude maker accused Alibaba’s Qwen team of orchestrating what it called the largest known ‘distillation attack’ against the company. According to Anthropic, Alibaba generated 28.8 million exchanges with Claude using nearly 25,000 fraudulent accounts over six weeks to improve its own AI models.

However, AI researcher Nathan Lambert argues that while frontier models are used for distillation, their role is often exaggerated. “The part where this SFT distillation is likely very useful is when a model moves into a new domain,” he wrote on X, adding that supervised fine-tuning (SFT) is only “the start of a long, strenuous process” that involves generating more training data, filtering it, and extensive reinforcement learning.

Lambert also said it was unlikely that Anthropic’s Fable 5 played a meaningful role in training recent Chinese models such as Kimi K3 and GLM 5.2. 

According to Lambert, as AI labs increasingly rely on reinforcement learning and multi-teacher training techniques, the relative importance of distillation is declining. “Altogether, the evidence in post-training is that distillation is becoming less impactful,” he said, adding that strong open-weight models could themselves become valuable teachers for future AI systems.

Meanwhile, Xiaoyin Qu, Founder of AI startup Tycoon.us, remarked that Anthropic’s proposed policies would not deter Chinese AI firms. “Kimi will continue to distil Anthropic and launch better models,” Qu wrote on X, adding that banning Kimi in the US would “only prove their tech works and help with their China IPO.”

She also said that Anthropic’s intellectual property stance could backfire by strengthening similar copyright claims against AI companies from publishers, authors, and the entertainment industry over the use of copyrighted material to train AI models. Notably, Anthropic recently agreed to pay $1.5 billion to settle a class-action copyright lawsuit brought by authors and publishers over its use of pirated books to train Claude. 

David Sacks, Co-chair of the President’s Council of Advisors on Science and Technology, in his recent appearance on All In Podcast, claimed that Anthropic was attempting to panic the government into banning Chinese open-source models to protect its own business interests and stifle competition.

Sacks explained that distillation is a standard industry practice, not IP theft, and any such ban would affect the US developer ecosystem. He argued that banning open-source models would harm the American developer ecosystem. 

“I think it would be a tragic mistake if the government were to take action against the open-source ecosystem. That would do nothing but hurt America's position in this AI race. It would backfire badly,” he said.