Chinese AI startup Moonshot AI publicly released the Kimi K3 model weights and key infrastructure tools on July 27, 2026, allowing developers worldwide to download, customize, and distribute the model freely [1, 2, 3, 4, 5].
With 2.8 trillion parameters, Kimi K3 is now the largest open-weight AI model globally [1, 3, 5]. It employs a Mixture-of-Experts (MoE) architecture and uses the MXFP4 format. The full model occupies about 1.4 terabytes of storage and requires significant computational resources — around eight servers each equipped with eight Nvidia H100 or B200 GPUs (64 total accelerators) — to run efficiently [5].
The Kimi K3 model supports a massive one-million-token context window and offers an OpenAI-compatible API. This enables easy integration for teams already using OpenAI or Anthropic models [5]. Moonshot’s ecosystem has expanded beyond Nvidia hardware as Alibaba’s cloud supernode with Zhenwu M890 chips and Huawei’s Ascend 910C platform have adapted to support Kimi K3 [3].
Founder Yang Zhilin said Moonshot aims to attract users by prioritizing openness and broader availability compared to proprietary US AI systems [1]. Since announcing Kimi K3, Moonshot’s daily sales have surged at least sixfold as the company seeks a new funding round valuing it around US$50 billion ahead of a potential IPO in Hong Kong [1, 4]. Bloomberg Intelligence analysts Mandeep Singh and William Tong noted that adoption of open-weight large language models is accelerating, but successful monetization will depend on distribution through large cloud providers [4].
The release of Kimi K3 has sparked heightened discussions in Silicon Valley amid concerns over Chinese advances in AI, with calls for bans on Chinese open-source AI models intensifying [3, 4].
