Moonshot AI releases Kimi K3 as China narrows gap with U.S. models

Moonshot AI has released Kimi K3, a new version of its Kimi large language model that the company says can compete with leading U.S. systems. The launch adds pressure to Anthropic and OpenAI as businesses weigh the cost of using top-tier AI models.

The Chinese startup unveiled Kimi K3 on July 16. Moonshot says the model has 2.7 trillion parameters, a measure of the internal weights that help determine how much complexity a model can handle; Fortune reported that makes it the largest open-weight large language model currently available. DeepSeek V4 has 1.6 trillion parameters, according to Fortune.

Moonshot described K3 in its release as its strongest open-source coding model so far. The company said the model can run extended engineering tasks with limited human supervision, work across large codebases and coordinate terminal tools.

Benchmark claims put K3 near U.S. leaders

Moonshot said K3 performs competitively with Anthropic’s Fable 5, which Fortune described as the most advanced widely available AI model on the market. Moonshot also said K3 beat Anthropic’s Opus 4.8 and OpenAI’s GPT 5.6 Sol and GPT 5.5 by a wide margin on its own tests.

On benchmarks released by Moonshot, K3 regularly appears among the top three models. If those results hold up under outside testing, Fortune reported, K3 would be one of the clearest signs that Chinese developers can produce open-weight systems in the same tier as the strongest models from Anthropic and OpenAI.

Anthropic’s Mythos 5, the model family behind Fable 5, has been reported by Fortune to be the most capable known model for cyber-related tasks. Access to Mythos is limited to companies in Anthropic’s Glasswing program, which Fortune reported was created to help critical-infrastructure organizations find and fix software vulnerabilities.

Policy debate grows around Chinese models

Fortune reported that analysts had not expected China to produce a model as capable as Fable until early next year. K3’s release could sharpen the U.S. debate over whether controls on frontier AI should be loosened to help American companies compete or tightened to slow China’s AI sector.

The U.S. government temporarily placed export controls on Mythos and Fable after Amazon researchers found a way to bypass Fable’s safeguards and expose Mythos’ cyber capabilities, according to Fortune. U.S. officials also initially instructed OpenAI to restrict GPT-5.6 access to selected trusted partners, Fortune reported.

Bloomberg has reported that U.S. lawmakers are considering penalties aimed at stopping Chinese firms from using distillation, a method in which the outputs of a stronger model help train a smaller or more efficient one. Anthropic has accused Moonshot, z.ai, Minimax, Alibaba and DeepSeek of illicit distillation attacks, according to Fortune. CNBC has reported that U.S. officials are also discussing ways to reduce the appeal of Chinese open-source models, including by backing U.S. open-source alternatives.

Chinese AI models have gained users because they tend to cost less and can be adapted by developers who download open-source versions, according to Fortune. Companies using them often need more technical skill and must rent cloud-based AI chips to host the models.

U.S. export controls have limited Chinese developers’ access to advanced AI processors, Bloomberg reported. Yutong Zhang, Moonshot AI’s president, said at the World Economic Forum earlier this year that the company lacked the option of scaling compute freely and focused instead on research and efficiency.

Moonshot’s business grows

Moonshot’s earlier Kimi models had already found users in Silicon Valley. Fortune reported that Cursor used Kimi while building Composer 2, and DoorDash chief technology officer Andy Fang said in an early July social media post that the company assigns lower-level work to Kimi K2.6. Thinking Machines used Kimi K2.5 to create early post-training data for its Inkling model, according to Fortune.

K3 costs $15 per million output tokens, according to Fortune. That is higher than z.ai’s GLM-5.2 at $4.40 and DeepSeek V4 at $0.87, but below Fable’s $50 price for the same amount of output.

Bloomberg reported in May that Moonshot raised $2 billion at a valuation above $20 billion. Fortune said a statement from the company’s financial adviser put Moonshot’s annual recurring revenue above $200 million and listed Alibaba, Tencent, Meituan and Hongshan Capital among its backers.

This story draws on original reporting from Fortune.