Chinese startup Moonshot AI has unveiled a new model it says closes the gap with the best US offerings and surpasses the more capable OpenAI and Anthropic systems in some benchmarks.
Kimi K3 still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol in overall performance, the company said Friday, but it consistently outperforms other tested models.
The model beats Claude Opus 4.8 and GPT 5.5 — models that sit just behind the best Anthropic and OpenAI programs — in benchmarks that include generic codes and agents, according to Moonshot.
It is China’s largest AI model to date, with 2.8 billion parameters, referring to the size of its neural network.
“Despite ongoing hardware/compute capacity issues in China, K3 shows that pre-training scaling, coupled with architectural innovation, can still bring benefits to replace China’s leading models,” Bank of America analysts said in a paper led by Alex Liu.
The release comes as the race for AI supremacy between the US and China intensifies.
Chinese AI models are already gaining traction among Western companies as they close the performance gap with US rivals and remain cheaper to use than the more advanced offerings from American labs. US lawmakers are considering how to curb the growing adoption of Chinese AI models by domestic companies.
Another DeepSeek season?
Patrick Moorhead, CEO and senior analyst at Moor Insights and Strategy, noted the market’s reaction to the new Kimi K3 model as “a shock similar to the DeepSeek shock,” explaining in a post on X that despite technological advances, “We are far from brilliant.”
Moorhead said in a post that large-scale linguistic models, or LLMs, like Kimi K3 will only “accelerate and expand the thinking market faster than the outside,” emphasizing the general shift in the technology field from focusing only on the size and conceptual power of the model itself to the general application that the technology allows.
Perplexity CEO Aravind Srinivas told CNBC last week that the focus for startups and developers is on finding the best ways to use AI models to power their applications, rather than focusing entirely on one big, fundamental system. That’s part of the reason why OpenClaw’s freely available technology became so popular with developers earlier this year; so-called harnesses allow coders to easily exchange and generate different AI models that empower digital assistants to take a series of actions without needing to rely on a single LLM.
“A model alone is no longer a product,” Srinivas said at the time. “Harness, a singing system that puts the model inside a very capable harness and pairs the model with a lot of tools.”
Moorhead says what he believes is an overreaction to the Kimi K3 release is political, telling CNBC in an email that “There is a big debate in Washington DC about whether the US should use Chinese open source models and whether US companies should allow the Chinese to use their models.”
“What’s next is strange as the Chinese seem to be doing well with their model,” Moorhead said.
Lu Zhang, founder and managing partner of Fusion Fund, said that despite the widespread attention models like Kimi K3 can get, most of the developers who use the technology “come from the startup ecosystem, a little bit from the big corporate side.” These codes will often replace one AI model if there is a more powerful version available or at least a cheaper and more efficient one to use in their respective applications, he explained.
And while these types of AI may seem extremely powerful at first, they are not “plug and play” and require a lot of technical knowledge from developers to use their basic capabilities, Zhang said.
Although general discourse involving the open-source AI model space often involves “the broader issue of US-China competition,” Zhang said there are several US companies that are increasingly releasing open-source AI models, including Thinking Machines and DeepReinforce, which is backed by his venture capital fund.
It’s only a matter of time before a more advanced open-source AI model captures the zeitgeist, given how fast the AI space is moving, he said.
Similar to how DeepSeek’s R1 AI model in 2025 generated attention for its potential cost-effectiveness compared to fitness technology, the current hoopla over Kimi K3 can be attributed to rising concerns about AI’s overall cost and ability to generate a return on investment.
Simon Koser, chief product officer at AI startup Tzafon, said the Kimi K3 is legitimately impressive because it excels in areas such as coding, and AI labs developers can find compelling.
“Cost has become a big issue for some of these labs,” Koser said, stressing that AI leaders like Anthropic and OpenAI may feel pressured by cheaper AI models available on the market.
However, there are many ways to use the technology, and not every AI model is the best for every task despite what early benchmark tests may show. Certain AI models may react differently when put into production when tested, and there is no jack-of-all-trades AI model that is superior to everything else on the market.
“It’s going to look like a lot of people are changing,” Koser said. “But in practice, I’m not sure the change is that big.”
China’s AI shock
Founded in 2023, Beijing-based Moonshot AI is one of the leading model makers in China. It raised $2 billion in a $20 billion-plus deal in May, Bloomberg reported.
Backers include Chinese tech giants Alibaba, which makes the Qwen series of AI models, and Tencent.
Shares of China’s AI rivals fell on news of the release. Z.aiwhich released a new model to much fanfare in June, its stock fell 28% on Friday. MiniMax Groupanother Chinese model company, down 16%.
“K3 raises the ceiling of China’s AI models, shifting the burden of proof to other private AI labs,” Liu said.
Earlier this week, Alibaba he saw his stock being raised by the news he was dealing with an apple in China. However, shares fell 4% on Friday.
“For Alibaba, although it benefits from extensive AI training/growth in the use of its cloud resource provided by a strong computing environment, Alibaba Qwen’s “open source leader” narrative may face some tests,” Liu said.
