Chinese AI reaches new creative heights
Moonshot AI, a Chinese startup, has released its latest large language model, Kimi K3, closing the performance gap with U.S. AI leaders while businesses evaluate the high costs of models from Anthropic and OpenAI.
The new model, announced on July 16, contains 2.7 trillion parameters, making it the largest open-weight LLM available. Parameters determine a model’s internal complexity, with higher counts generally enabling more advanced reasoning. DeepSeek V4, by comparison, has 1.6 trillion parameters.
On benchmarks, K3 performs competitively with top models
Moonshot stated Kimi K3 performs at a level similar to Anthropic’s Fable 5, currently the most advanced widely available AI model. The company also reported that its model surpassed Anthropic’s Opus 4.8 and OpenAI’s GPT 5.6 Sol and GPT 5.5 in internal tests, consistently ranking within the top three models.
Fable 5 derives from Anthropic’s Mythos 5, a restricted model available only to select enterprises through the company’s Glasswing program. This initiative assists critical infrastructure providers in addressing software vulnerabilities. Mythos 5 is regarded as the most capable model for cybersecurity tasks, though access remains limited.
If K3’s performance claims prove accurate, the release demonstrates that Chinese developers can build open-weight systems matching U.S. counterparts. Analysts had not anticipated China producing a model at Fable’s level until early next year.
The timing coincides with U.S. policy debates over frontier AI regulation. The government temporarily restricted exports of Mythos and Fable after researchers bypassed Fable’s safeguards, exposing Mythos’ cyber capabilities. OpenAI also faced initial pressure to limit GPT-5.6 releases to approved partners.
K3’s early launch may influence these discussions. Some policymakers could argue for relaxed controls to maintain U.S. leadership, while others might advocate stricter measures to slow China’s AI progress. The release also intensifies scrutiny of distillation practices, where smaller models train on outputs from larger ones. Anthropic has accused Moonshot, z.ai, Minimax, Alibaba, and DeepSeek of unauthorized distillation, prompting U.S. officials to explore ways to reduce the appeal of Chinese open-source alternatives.
Cost and flexibility drive adoption of Chinese models
Chinese AI models are gaining global users due to lower pricing and open-source adaptability. Developers can freely download and modify these systems, though deploying them often demands more technical skill and cloud-based AI chips.
U.S. export restrictions have blocked Chinese firms from accessing advanced AI processors, pushing them to prioritize efficiency. “We couldn’t rely on scaling up compute,” said Yutong Zhang, Moonshot’s president, at the World Economic Forum earlier this year. “That forced us to focus on fundamental research and optimization.”
This isn’t the first instance of Chinese AI exceeding expectations. In 2023, DeepSeek released a model that outperformed Meta’s Llama 2 despite having fewer parameters, proving efficiency gains could compensate for hardware limitations. K3’s launch indicates this trend is accelerating, with Chinese developers now competing at the highest levels of AI performance.
Moonshot secured $2 billion in funding during May, bringing its valuation to over $20 billion. Annual recurring revenue surpasses $200 million, according to its financial advisor. Investors include Alibaba, Tencent, Meituan, and HSG, formerly Sequoia China.
The company plans an IPO in Hong Kong, following peers MiniMax and z.ai, which went public in January. DeepSeek is considering a listing in Shanghai.
Lower costs and open-source flexibility have made Chinese models attractive to global users. Business leaders increasingly weigh these factors when choosing AI solutions.
