Source: RFI
What makes China's Kimi K3 so remarkable? "It has become the only topic of conversation in Silicon Valley."
On Monday, July 20, Thomas Pontiroli, a reporter for the Science section of Le Monde, reported that Moonshot AI shook the entire artificial intelligence community by releasing the largest open-source AI model ever.
Although the "DeepSeek moment" has not yet been replicated, this startup has indeed narrowed the gap between China and the U.S. in AI to an unprecedented degree.
According to the UK AI Safety Institute, the gap between Chinese and U.S. models has narrowed to 4 to 7 months, down from 6 to 10 months last year.

This news shook U.S. AI labs over the weekend. On Thursday, ahead of the World Artificial Intelligence Conference in Shanghai, Chinese startup Moonshot AI unveiled its new flagship model: Kimi K3, a 2.8-trillion-parameter open-source model, the largest of its kind to date.
This Beijing-based startup is backed by Alibaba with a 36% stake. In May of this year, it raised $2 billion at a valuation exceeding $20 billion and is preparing for an IPO. Its founder, Yang ZhiLin, previously co-authored papers with Yann LeCun and Yoshua Bengio. Its earlier K2 model has already been adopted by U.S. software company Cursor and food delivery platform DoorDash. Thus, it is no obscure player—but this news still caught the entire industry off guard.
Concerning
The report noted that Box co-founder and CEO Aaron Levie told Bloomberg, “This is undoubtedly a surprise to many; it has become the only topic of conversation in Silicon Valley.” He considers it a “major advancement” for open-source models. David Sacks, who served as the White House’s AI lead in the Trump administration, called the development “concerning.”
Based on the only available Chinese company data currently, K3 significantly outperforms Anthropic’s benchmark model Claude Fable 5 in long-form programming tasks (42 vs. 35) and online information retrieval (91.2 vs. 88), but still lags in traditional software development (81.2 vs. 86.6).
The report noted that the results were closer to OpenAI’s latest top model, GPT-5.6 Sol: nearly tied in scientific reasoning (93.5 vs. 94.1). Moonshot AI also acknowledged a “clear gap” compared to U.S. AI in user experience. Nevertheless, certain sectors on Wall Street experienced volatility on Friday.
K3 effect
The report noted that on Friday the 17th, NVIDIA fell 3%, AMD dropped as much as 5%, Intel declined 4%, and Applied Materials and Lam Research both fell more than 4%. The Philadelphia Semiconductor Index slid as much as 5.7% (before recovering somewhat). The index has declined over 20% since its late-June high, and the global semiconductor industry has lost $3.3 trillion in market value since June 22.
The impact of K3 also reached the Chinese market itself. On the same day, the stock prices of two companies listed in Hong Kong since January, which directly compete with Moonshot AI, plummeted: Zhipu AI, which had been highly favored just two weeks ago for its GLM-5.2 model, dropped by as much as 28%, while MiniMax fell by 16%.
Some are viewing Kimi K3 as the new "DeepSeek moment." In January 2025, DeepSeek released the R1 model, which not only matched the performance of contemporary ChatGPT and Claude but was also free and significantly cheaper to train. This move shook the entire AI investment logic at the time—industry giants that had invested billions of dollars were disrupted; NVIDIA, the biggest beneficiary, saw $590 billion in market value evaporate in a single day.
The market is "overreacting"
The report analyzes that, however, this time the situation may be different. First, it is still unclear under what terms Kimi K3 will be made available: Moonshot AI has promised to open access on July 27, but has not yet released any licensing agreements. Additionally, China may also impose export controls on this industry, as it has in other sectors.
Secondly, given today’s increasing sensitivity to computing costs and opposition to new data center construction, this model is overly large. On-premises deployment of K3 requires at least 64 processors, and its price has tripled compared to the previous generation. Developer Simon Willison noted: “This is the most expensive model commercialized by a Chinese AI lab to date.” In an environment where enterprises are increasingly focused on the profitability of AI projects, this could also become a barrier.
The report also noted that some analysts believe the market reaction has been excessive. Patrick Moorhead of Moor Insights compared it to the panic surrounding DeepSeek, while Bernstein viewed it as a normal catch-up by China’s technological capabilities. Just like with DeepSeek, “not surprising” does not mean “not important.”
Former Trump "AI Czar" speaks out
The report quoted Ion Stoica, co-founder of Databricks and founder of the AI model ranking platform Arena, telling Bloomberg: “Previously, people thought Chinese open-source models lagged behind the most advanced models by 6 to 9 months; now that gap has narrowed to 2 to 3 months.” In the field of cybersecurity, the UK’s AI Safety Institute estimates that the gap between Chinese and U.S. models has narrowed to 4 to 7 months, down from 6 to 10 months last year.
In July this year, OpenAI CEO Sam Altman told U.S. media CNBC: "China's open-source models are becoming increasingly strong." However, he still believes the company can maintain a leading edge on its most advanced models. Anthropic continues to assess that the lead could last up to a year.
The report concludes with a warning, but in David Sacks’ view, the catch-up is indeed happening—and at a bad time: after Anthropic restricted access to Claude Opus 5 to non-U.S. users, CNBC reported that the Trump administration would continue deciding which companies can access the latest models (though this claim has been denied). To this former “AI czar,” this is a barrier, and he bluntly states: “This is how we lose the AI race!”
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush
