Since late 2022, US export rules have prevented Nvidia from shipping its most advanced chips to China. This was supposed to slow China's AI progress. On July 17, 2026, the Beijing-based company Moonshot AI unveiled the world's largest freely available language model: Kimi K3 with 2.8 trillion parameters. Investors responded with a selloff that gave the Philadelphia Semiconductor Index its worst trading week in more than 15 months.
The world's largest open language model
Moonshot AI was founded in March 2023 by Yang Zhilin, together with Zhou Xinyu and Wu Yuxin. Yang studied at Tsinghua University in Beijing, earned his doctorate at Carnegie Mellon University, and conducted research at Google Brain and Meta during his PhD. The company was valued at over 20 billion dollars in its last funding round. According to a Bloomberg report from July 19, Moonshot AI is now exploring a Hong Kong IPO with a target valuation of around 30 billion dollars.
Kimi K3 uses a Mixture-of-Experts architecture that the company calls "LatentMoE." Of 896 specialized expert networks, only 16 are activated per token, corresponding to an activation rate of 1.8 percent. With 2.8 trillion total parameters, a single query effectively engages around 50 billion parameters. This design allows high overall capacity with lower computational demand per query. The context window is one million tokens; images, text, and video are processed natively.
Starting July 16, Kimi K3 has been accessible via the Kimi platform and API. Prices stand at 0.30 dollars per million cache-hit input tokens and 15 dollars per million output tokens. On July 27, 2026, the full model weights are scheduled for free release under a modified MIT license.
Stronger in coding, not across the board
In the Frontend Code Arena, a programming competition based on anonymized developer judgments, Kimi K3 achieved 1,679 points according to Moonshot AI and ranked first ahead of Claude Fable 5. In coding and agent benchmarks, the model reportedly exceeds Claude Opus 4.8 and GPT 5.5. In general performance comparisons, Kimi K3 trails Claude Fable 5 and OpenAI's GPT 5.6 Sol, currently the most powerful US models.
Bernstein analyst Robin Zhu calls the release "confirmatory" of the thesis that China's AI sector is keeping pace with global development. However, he cautions that convergence in reasoning capabilities at the frontier is "directionally negative for the long-term margins of AI model labs." If multiple models deliver similar performance, pricing power for Anthropic and OpenAI declines.
It should be noted that benchmark rankings are based on Moonshot's own claims and have not yet been reproduced by independent third parties. Benchmark reports from model developers tend toward favorable self-presentation. Developer communities worldwide will be able to independently verify results after the weight release on July 27.
The second DeepSeek moment in 18 months
The market reality of the past week strongly resembles the DeepSeek-R1 shock of January 2025. Back then, reports of a Chinese open-source model cost Nvidia roughly 600 billion dollars in market value in a single trading day. Kimi K3 repeated this effect and stretched it across an entire week: since mid-June chip sector highs, the segment has lost around 3.3 trillion dollars combined. Nvidia fell 2.2 percent on July 18, Applied Materials lost 5.6 percent.
What makes the investor reaction particularly revealing is the logic behind it. US chip stocks have become the barometer for the AI growth thesis: more AI demand means more compute demand, means more chip sales. Kimi K3 challenges this causal chain. If a model operates with 1.8 percent active parameter utilization per query and is also made available for free, it is unclear how many of the planned data centers with expensive Nvidia chips are actually needed.
Especially symbolic is the geopolitical finding: Kimi K3 emerged despite US export restrictions that prevent Nvidia from supplying its most advanced chips to China. The company has not disclosed exactly how it organized model training at this scale. Additionally, the sector came under pressure after reports that Kimi K3 designed a chip in 48 hours without proprietary design software. Following this, leading chip design software providers Cadence and Synopsys also declined sharply.
Weights released on July 27
The crucial date is July 27, 2026. On that day, the full model weights are scheduled for free download. One caveat: Kimi K3 with 2.8 trillion parameters cannot run on consumer hardware. Local operation requires compute clusters with dozens of high-performance GPUs, limiting deployment to well-equipped research institutions, technology companies, and cloud providers. Nevertheless, the opening remains significant: anyone can study the architecture, adapt it, and further develop it without accepting a vendor's terms of service.
Robin Zhu from Bernstein warns that Washington will watch the situation closely: "It would not surprise us if K3 accelerated Anthropic's campaign for regulatory influence against Chinese competitors." After the DeepSeek shock in January 2025, the US Congress discussed whether open weights from Chinese models should be restricted. Concrete measures were not taken. With Kimi K3, the question is back on the agenda: whether the US wants to restrict downloading Chinese open-weight models and whether such a measure would be technically enforceable.