Moonshot Releases Kimi K3, the Largest Open AI Model Ever
Moonshot AI made Kimi K3's full weights free to download, giving developers a near-frontier model to self-host and putting new pricing pressure on US AI labs.
On this page
Moonshot AI released the full weights of its Kimi K3 model for public download on Monday, making the 2.8-trillion-parameter system the largest open-weight AI model ever offered to developers and sharply escalating the rivalry between Chinese labs and American frontier developers, according to Bloomberg.
The release means anyone can download, modify, and self-host a model that independent evaluations place within striking distance of the best proprietary systems from Anthropic and OpenAI. Founder Yang Zhilin has said the Beijing-based company wants to grow its user base through openness and broader availability than the closed US systems it competes against, Bloomberg reported. A technical report detailing the model's architecture, training, and benchmark performance is expected to follow the weight release.
A frontier-class model, free to download
Kimi K3 debuted as a hosted service earlier this month and immediately reshuffled expectations for open models. The system supports a context window of one million tokens, enough to feed an entire large codebase or a stack of lengthy documents into a single query. In blind evaluations on Arena, developers selected K3 ahead of leading US models on front-end coding tasks. Overall it placed below Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, but it outperformed both companies' prior-generation models across coding and general agent evaluations.
Moonshot itself acknowledged the gap at the top while claiming the model "demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models." Independent analyses from Arena.ai and Vals AI broadly supported the claim that K3 is competitive with flagship frontier systems, TechCrunch reported.
The scale is unprecedented for an open release. At 2.8 trillion parameters, K3 towers over previous open-weight record holders from Chinese rivals, and its arrival extends a pattern that began when DeepSeek released its R1 model in January 2025 and briefly wiped hundreds of billions of dollars off US chip stocks.
Markets and rivals felt the debut
K3's launch two weeks ago rattled both Chinese and American markets. Shares of Chinese AI competitor Z.ai lost as much as 30 percent of their value in Hong Kong trading, MiniMax Group fell as much as 16 percent, and Alibaba dropped 4 percent. In the US, the Nasdaq slipped about 1 percent as investors sold chipmakers, a reaction amplified by the launch's timing alongside a speech from Chinese president Xi Jinping at the World AI Conference in Shanghai.
The model has also transformed Moonshot's business. The company's daily revenue has grown by a factor of at least six since K3 launched, according to Bloomberg. Moonshot reached $300 million in annual recurring revenue in June, up from $200 million in April, and is reportedly seeking a new funding round at a $50 billion valuation ahead of a potential Hong Kong initial public offering as soon as this year.
Washington pushes back
The open release lands in the middle of an intensifying policy fight. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot last week of training K3 on banned Nvidia chips and of conducting large-scale distillation against US models, including Anthropic's Fable model. Moonshot has not responded to the allegations.
Reaction across the US tech industry has split along now-familiar lines. David Sacks, co-chair of the President's Council of Advisors on Science and Technology, pointed to K3 as evidence that US regulation is slowing domestic developers. Former Uber chief executive Travis Kalanick echoed complaints about Chinese labs distilling from American models. OpenAI's head of strategic futures Dean Ball called Kimi "a very good model" whose performance likely cannot be explained away by distillation, while warning that a world dominated by open-weight models carries its own risks and predicting that regulators will eventually create compliance pressure around Chinese open models in regulated industries.
What it means for founders and operators
For startups, the immediate effect is pricing leverage. A near-frontier model that any hosting provider can serve puts downward pressure on API costs across the board, and gives teams with sensitive data a realistic self-hosting path that did not exist at this capability level before. Companies in healthcare, legal, financial services, and government-adjacent work can now run a top-tier model inside their own infrastructure instead of sending data to a third-party API.
The catch is operational and regulatory. Running a 2.8-trillion-parameter model is not a laptop project; practical deployment requires serious GPU infrastructure or a specialized hosting provider. Notably, none of the three major hyperscale cloud platforms, Amazon's AWS Bedrock, Microsoft's Azure Foundry, or Google's Vertex AI, currently offer Kimi K3 or comparable Chinese open-weight models, Bloomberg Intelligence analysts Mandeep Singh and William Tong noted. That leaves an opening for smaller inference providers that specialize in hosting third-party open models, and those companies are expected to be the release's most direct beneficiaries.
Founders selling into regulated enterprises should also weigh the policy risk. If federal agencies begin issuing guidance that casts doubt on Chinese-origin models, as some in the industry predict, products built on K3 could face procurement friction that products built on US models will not. The prudent move for most teams is to keep model choice abstracted behind an interface layer so a swap remains cheap if the compliance picture shifts.
The bigger signal is strategic. Moonshot is betting that openness, not exclusivity, is the fastest route to becoming the default platform for the world's developers. With six-fold revenue growth since launch and a $50 billion valuation round reportedly in motion, the bet is working so far. American labs now face an open-source competitor operating at a scale DeepSeek never reached, and the response from Washington and Silicon Valley over the next few months will shape how freely startups everywhere can build on it.
Frequently asked questions
What exactly did Moonshot AI release on July 27?
Moonshot released the full weights of Kimi K3 for unrestricted public download. Developers can deploy, adapt, and self-host the 2.8-trillion-parameter model rather than accessing it only through Moonshot's hosted API. A technical report on the model's design and training is expected to follow.
How good is Kimi K3 compared with GPT-5.6 Sol and Claude Fable 5?
Independent rankings place K3 below Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol overall, but ahead of both companies' prior-generation models on coding and agent evaluations. In blind Arena tests, developers picked K3 over leading US models on front-end coding tasks.
Can a startup actually self-host a 2.8-trillion-parameter model?
Only with substantial GPU infrastructure, which puts direct self-hosting out of reach for most early-stage teams. The more practical route is running K3 through inference providers that specialize in hosting open-weight models, since the major hyperscale clouds do not currently offer it.
Why is the US government concerned about Kimi K3?
The White House Office of Science and Technology Policy has accused Moonshot of training K3 on export-restricted Nvidia chips and of distilling from US models, including Anthropic's Fable. Industry figures have also predicted regulators may discourage use of Chinese open models in regulated sectors.
What should founders building on AI do about this release?
Treat it as pricing leverage and a fallback option. Even teams that stay on US frontier APIs benefit from the competitive pressure on cost. Keeping model choice abstracted behind an interface layer preserves the option to switch if pricing, capability, or compliance conditions change.
Sources
The Briefing
Practical business intelligence for people building in the AI era.
One concise email with the reporting, guides, and tools that matter to founders and operators. No noise, no spam, unsubscribe anytime.
