Chinese startup Moonshot AI released Kimi K3 on July 16, a 2.8 trillion-parameter model that is now the largest open-weight AI system in the world. Backed by Alibaba, Tencent and Meituan, the Beijing lab has built a system whose benchmark results have outperformed the leading closed U.S. models, including Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, trailing only the newest Claude Fable 5 and GPT-5.6 Sol.
The model arrives as the third generation of Moonshot’s Kimi line, which began as a chatbot in 2023 and built its reputation on the open-weight K2 family. K2, released in July 2025, was already known for strong performances on coding benchmarks; its successor K2.6 followed in April 2026.
K3 dwarfs both its Chinese rivals and its own predecessor: it is roughly 75% larger than DeepSeek’s V4 Pro, at 1.6 trillion parameters, and nearly four times the size of Zhipu AI’s GLM 5 series, at 744 billion parameters. Two versions shipped at launch, K3 Max for chat and agent tasks and K3 Swarm Max for large-scale parallel processing, both available first through Kimi Code and the Kimi app.
Kimi K3 is a mixture-of-experts model, meaning only a fraction of its parameters activate on any given request. Specifically, it draws on just 16 of its 896 experts per token, roughly 1.8% of the total pool, which keeps inference costs down despite the model’s size.
It ships with a 1-million-token context window and native visual understanding, and it is built on two architectural innovations developed internally at Moonshot: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a drop-in replacement for residual connections. Both techniques had already been published as open research by the Moonshot team on GitHub before the model’s release.
On Artificial Analysis’s private long-horizon knowledge work evaluation, K3 reached an Elo of 1,547, a 732-point jump from its predecessor, K2.6, placing it behind only Claude Fable 5. The model also took the top spot on LMArena’s Frontend Code Arena, surpassing Fable 5 in that specific benchmark. K3 is roughly 2.8 times the size of K2.6, yet Moonshot says it uses 21% fewer output tokens than its predecessor on equivalent tasks, pointing to efficiency gains that go beyond simply adding parameters.
AI safety researcher Ryan Greenblatt placed the model in the same tier as Opus 4.8, though he described it as “somewhat more benchmaxxed,” a caveat that reflects the broader difficulty of comparing models trained partly to excel at the specific evaluations labs use to promote them.
Independent developer Simon Willison, who tested K3 the day of release, noted that the model currently offers only a single reasoning setting, “max,” which drove up token usage in his tests: a simple request to generate an SVG illustration consumed over 16,000 output tokens, most of them spent on internal reasoning, for a cost of roughly 25 cents.
Weeks, not months: how far China has closed the gap
The clearest summary of what this release means came from Ryan Fedasiuk, a researcher at the American Enterprise Institute, in a report published under the title “China has caught up to the U.S. in frontier AI.” Fedasiuk’s conclusion, cited by Xataka, is that the gap between the two countries, previously measured in six to eight months, has narrowed to something closer to a few weeks.
K3’s pricing reflects that shift. The API runs at $3 per million input tokens and $15 per million output tokens, the highest rate of any Chinese lab to date and roughly in line with Anthropic’s Sonnet-tier pricing, a sharp increase from the $0.95/$4 pricing of K2.6. Even so, Artificial Analysis puts the real-world cost per task at $0.94, close to GPT-5.6 Sol’s $1.04 and about half of Opus 4.8’s $1.80. Full open-weight release is scheduled for July 27, after which the model can be self-hosted without ongoing API costs.
The launch itself began with a leak: a promotional page on Moonshot’s Kimi Open Platform revealed the model a day before the company’s official announcement on July 16, when K3 went live on Kimi Code and inside the Kimi app. The timing placed the release just ahead of the 2026 World Artificial Intelligence Conference in Shanghai.
Moonshot’s financial position has grown alongside the model. The company raised $2 billion at a $20 billion valuation in May and is reportedly now in talks for a new round that could value it at $30 billion, according to MLQ News, while domestic rival DeepSeek is already valued considerably higher.
Markets reacted quickly to K3’s release: the Nasdaq fell around 1% on July 17, with AI chip stocks including Nvidia declining as investors weighed the competitive threat. The sharpest losses, however, landed on Moonshot’s own Chinese rivals: Hong Kong-listed Zhipu AI dropped 28.49% that day, and MiniMax fell 15.62%, as the market priced in the risk that K3 would draw developers and enterprise contracts away from smaller domestic labs.
Washington weighs a response
The release has also revived a policy debate in Washington that predates K3 itself. The National Security Agency and the White House Office of the National Cyber Director previously considered issuing a security warning telling U.S. companies not to use Chinese AI models, a push that stalled after pushback from officials including former White House senior policy adviser on AI Sriram Krishnan.
According to an Axios report cited by multiple outlets, the Trump administration is now reconsidering those restrictions, citing cybersecurity concerns, following K3’s launch. Officials are reportedly weighing several tools rather than a single outright ban: adding Chinese AI labs to the Commerce Department’s Entity List, federal procurement restrictions, cybersecurity advisories, and liability rules that would place legal risk on U.S. companies using Chinese models.
Advisories from bodies like the Cybersecurity and Infrastructure Security Agency would not be legally binding, but they could still create liability exposure for companies that continue using flagged software.
The position is not unanimous within the administration. David Sacks, chair of the President’s Council of Advisors on Science and Technology, has reportedly pushed back against broad restrictions, warning that they could undermine U.S. competitiveness rather than protect it. The debate comes amid other signs of instability in the administration’s AI regulatory apparatus, including the resignation of Chris Fall after three months as head of the Center for AI Standards and Innovation.
Xataka’s own reporting is careful to note that the absolute frontier in AI remains with U.S. labs: Claude Fable 5 and GPT-5.6 Sol are still ahead of every other model on the market, Kimi K3 included. What has changed, according to Fedasiuk’s report, is not who leads, but how quickly that lead is shrinking.
Once K3’s full weights are released on July 27, developers will be able to build directly on top of it without paying Moonshot ongoing licensing fees, a different economic model from the subscription and API-based approach used by OpenAI and Anthropic. That contrast, between an open-weight model closing in on the frontier and a business model built around keeping the frontier closed, is what analysts like Fedasiuk say Silicon Valley now has to account for when planning its next release cycle.
So, what’s your take: is this the moment China truly caught up in the AI race, or is it too early to call? Let us know in the comments!