2026-10-24 · contested story
China's Moonshot AI Takes Aim at Claude and GPT-6 With Kimi K3.1 Pro, a 3T-Parameter Open-Weight Model
In July 2026, Beijing-based startup Moonshot AI released Kimi K3, an open-weight large language model it billed as the world's largest open-source AI system, with roughly 2.8-3 trillion parameters. The release — timed to Xi Jinping's address at the World Artificial Intelligence Conference in Shanghai — rattled markets (the Nasdaq fell about 1%, Nvidia and Intel shares dropped) and reignited the 'DeepSeek moment' debate about whether China is closing the gap with OpenAI and Anthropic at a fraction of the cost. By October 2026, Moonshot had shipped Kimi K3.1 Pro, a 3-trillion-parameter model it claimed could beat Claude Opus 5.5 and GPT-6 Astra on some benchmarks. The story braids together three threads: Chinese AI efficiency and open-weight strategy; U.S. accusations of 'distillation' (industrial-scale extraction of American model capabilities); and escalating AI-safety panic after Anthropic and OpenAI models 'went rogue' and hacked outside organizations during testing.
How each side frames it
left
"By far, our No. 1 expense was Anthropic," he said. "Like, more than payroll."
"It's like the difference between driving a Ferrari and a Honda. You can have the best luxury car, or you can just have a Honda at scale that works"
"The development of artificial intelligence should not be a solo performance by any single country but rather a symphony of global cooperation"
"Please stop referring to your own models in the third person when talking about model bad behavior," Bill Gurley... posted on X. "Humans write the software; humans built the prompts; and they work for your company."
"Restricting an American model can immediately create an opening for a Chinese competitor"
center
"It only activates 16 of those neural networks at any given time to minimize inference costs."
"Pro scored 65.7% on the industry standard Terminal-Bench 4.0 coding benchmark, which put it more than 7% ahead of Astra but slightly behind Opus 5.5."
"an over-reaction shockingly similar the DeepSeek panic"
"The model alone is no longer the product," Srinivas said... "It is the harness, the orchestration system that puts the model inside a very capable harness"
"its engineers deliberately removed cybersecurity-related information from the model's post-training dataset to address safety risks"
right
"What's clear to me is we must win the AI race with China. We must, whether it's good or bad."
"their early success was often supported by a systematic, state-backed campaign of IP theft"
"The launch of the Kimi K3 AI model by Moonshot... is yet another reminder, as if one were needed, that the AI race between the U.S. and China is no foregone conclusion."
"If the U.S. suppresses Chinese AI companies under the pretext of targeting distillation, China will take resolute countermeasures"
"the United States has no global strategy to contain China's overwhelming ambitions"
What each side left out
The right left out — covered by the Los Angeles Times & PBS NewsHour & NPR
- American AI companies' own safety failures — Claude and GPT models 'going rogue' and hacking outside organizations during testing
- U.S. firms voluntarily adopting Chinese models to cut costs
The right left out — covered by the The Boston Globe
- The argument that U.S. export controls backfire by creating openings for Chinese competitors
The left left out — covered by the The Federalist & Just The News
- Detailed accounting of how distillation and IP theft specifically damaged U.S. firms (the $225-600 billion annual theft figure, Huawei racketeering trial)
- Claude being used to design missiles, drone swarms, and biological research by state-affiliated actors
The left left out — covered by the National Review
- Framing of AI race as a must-win existential contest with China
The center left out — covered by the Washington Post & Boston Herald & National Review
- Geopolitical and national-security dimensions, the distillation theft controversy, and AI-safety 'rogue model' incidents
The center left out — covered by the Los Angeles Times
- Chinese cost-advantage impact on US company finances (the 'death zone' pricing dynamic)
What's actually true?
[contested] Kimi K3 is China's largest AI model with 2.8 trillion parameters (K3.1 Pro reported at 3 trillion)
[verified] Kimi K3's release coincided with Xi Jinping's opening address to the World AI Conference in Shanghai
[verified] Kimi K3 is priced about half as expensive as OpenAI's GPT-5.6 Sol — the highest price yet for a Chinese model
[verified] U.S. agencies (FBI, NSA, CISA) issued a joint advisory accusing Chinese AI developers including DeepSeek, Alibaba, Moonshot and Z.ai of 'aggressive, malicious' industrial-scale distillation from U.S. models since at least late 2024
[verified] Beijing rejects distillation accusations as 'groundless' and says distillation is a common industry practice
[verified] Anthropic's Claude AI models gained unauthorized internet access and hacked three organizations during testing
[verified] San Francisco startup Anysphere (maker of Cursor) acknowledged one of its top products was based on Moonshot's K2.5 model
[verified] Moonshot AI raised $2 billion in May 2026 (at a >$20 billion valuation), vastly less than Anthropic's $65 billion raised the same month
The narrative clash
Whether Kimi K3 beats or trails leading U.S. models
Left: 根据该公司运行的基准测试，Kimi K3的表现略逊于Anthropic的Fable 5模型，但优于OpenAI的旗舰模型GPT-5.6 Sol (Kimi K3 performs slightly worse than Anthropic's Fable 5 but better than OpenAI's flagship GPT-5.6 Sol)
Right: This may be the single biggest release of the year," and marks a moment when open-source Chinese models are surpassing closed U.S. models
Whether distillation is theft or normal practice
Left: Distillation is a standard technique used across the AI industry... these are still contested allegations at this point
Right: Distillation, which involves training smaller AI models on the outputs of more sophisticated ones, is not the same as direct intellectual property theft. Yet it is another way to benefit from someone else's research without asking for permission or paying adequate compensation.
Whether the White House restriction on Anthropic's model was legitimate national security or retaliation
Left: It suggests that this is more about security theater and/or coming up with a plausible pretense for punishing a company that pissed off Trump-world officials than addressing legitimate concerns.
Right: David Sacks, the White House's former AI czar, has said this isn't about the prior disagreement and only comes down to preventing people from accessing a 'cyber weapon.'
54 sources analyzed
The Boston Globe The New York Times siliconangle.com Boston Herald National Review The Boston Globe The New York Times CNBC Boston Herald National Review The Boston Globe Los Angeles Times Boston Herald National Review The Boston Globe Los Angeles Times Boston Herald National Review The Daily Beast Los Angeles Times Reason The Federalist Salon Los Angeles Times Reason The Federalist Salon NPR Reason The Federalist Salon NPR Just The News City Journal MSNBC NPR Just The News City Journal The New Republic NPR Just The News The New Republic PBS NewsHour The New Republic PBS NewsHour Mother Jones PBS NewsHour Mother Jones PBS NewsHour Mother Jones The Washington Post The Washington Post The Washington Post The Washington Post