Chinese startup Moonshot AI has released the latest version of its Kimi AI model, further shrinking the performance gap between Chinese and U.S. models just as global businesses are increasingly questioning the cost of deploying models from Anthropic and OpenAI.
On July 16, Moonshot AI unveiled Kimi K3, the latest version of its Kimi model. It boasts 2.7 trillion parameters, making it the largest open-weight large language model available today. (Parameters refer to the weights within the LLM; more parameters generally means that models can handle more complex reasoning.) DeepSeek V4 has 1.6 trillion parameters.
“K3 stands as Moonshot AI’s most powerful open-source coding model to date,” Moonshot AI wrote in a press release announcing the model’s release. “Operating with minimal human oversight, it can sustain long engineering sessions, navigate massive repositories, and orchestrate terminal tools.”
In its release, Moonshot claimed K3 performed “competitively” with Anthropic’s Fable 5, currently the most advanced AI model widely available on the market today, and “substantially outperformed” Anthropic’s Opus 4.8, and OpenAI’s GPT 5.6 Sol and GPT 5.5. On the company’s officially released benchmarks, K3 consistenty ranks within the top three models.
Anthropic’s Mythos 5 model, on which Fable 5 is based, is reportedly the most capable model in existence in performing cyber-related tasks, yet access to Mythos is restricted to a small number of enterprises that are part of Anthropic’s Glasswing program. That program was designed to help key makers of critical infrastructure to find and patch software vulnerabilities.
If K3’s performance claims hold, the model would mark one of the clearest signs yet that Chinese developers can build open‑weight systems in the same class as Anthropic’s and OpenAI’s, with direct consequences for global competition and a fast‑evolving debate over how to regulate frontier AI.
Analysts were not expecting China to produce a model as powerful as Fable until early next year.
K3’s release could also intensify discussions about the effectiveness of U.S. AI policy. The U.S. government temporarily imposed export controls on both Mythos and Fable after Amazon researchers found a way to jailbreak Fable’s guardrails and expose Mythos’ underlying cyber capabilities. It also initially told OpenAI to limit its release of GPT-5.6 to select trusted partners.
The revelation that a Chinese developer created a Mythos-level model months ahead of schedule could lead to looser controls in order to ensure the U.S. companies stay ahead–or it might invigorate hawks who wish to kneecap China’s AI sector as much as possible.
U.S. politicians are considering ways to stop Chinese developers from “distilling” U.S. AI models, which is when the outputs of a larger, more powerful AI model is used to help train smaller, more efficient models. Anthropic has accused Moonshot, z.ai, Minimax, Alibaba and DeepSeek of “illicit” distillation attacks. U.S. officials are also discussing ways to curb the appeal of open-source models from China, perhaps by encouraging the creation of U.S. open-source models.
Chinese AI models are winning converts around the world, due to their lower cost and greater efficiency. Also, as open-source models, developers can download the models for free and tweak them to suit their own purposes. But using open source models does often require more technical expertise on the part of the companies deploying them. It also requires those companies to rent AI chips through cloud providers in order to host the models.
U.S. export controls barred Chinese developers from getting access to the advanced AI processors used to train and run the most powerful AI models. That forced Chinese developers to find new ways to get more bang for their computing buck.
“We knew we didn’t have the luxury to simply scale up compute,” Yutong Zhang, president of Moonshot AI, said at the World Economic Forum earlier this year. “That forced us to focus on fundamental research and efficiency.”
Moonshot’s previous AI models were already making inroads into Silicon Valley. Cursor, the vibe-coding startup, used Kimi to help build Composer 2, its AI coding agent; Doordash also delegates “lower-level work to Kimi K2.6,” according to chief technology officer Andy Fang in an early July social media post.
Thinking Machines also tapped Kimi K2.5 to generate early post-training data for its new Inkling model, released on July 15.
K3 is expensive—by Chinese standards. K3 costs $15 per million output tokens, compared to $4.40 per million output tokens for z.ai’s GLM-5.2 and $0.87 for DeepSeek V4. Still, it’s cheaper than the equivalent U.S. models: Fable costs a whopping $50 for the same amount of output.
Moonshot AI raised $2 billion in funding in May, valuing the company at over $20 billion. A statement from the company’s financial advisor stated Moonshot’s annual recurring revenue exceeded $200 million. Moonshot’s backers include all of China’s largest tech firms—including Alibaba, Tencent and Meituan—as well as HSG (formerly Sequoia China).
Moonshot’s fellow AI developers, MiniMax and z.ai, went public in Hong Kong in early January. Moonshot AI, too, is reportedly preparing for an initial public offering in Hong Kong; DeepSeek, in contrast, is considering a listing in Shanghai.
