Moonshot AI Aims to Secure Additional High-End NVIDIA Chips for Training Its Next-Generation Kimi K4 Large-Scale Model
2 day ago / Read about 0 minute
Author:小编   

Following the launch of its open-source large-scale model, Kimi K3, which boasts a staggering 2.8 trillion parameters, Chinese AI startup Moonshot AI has once again found itself at the center of geopolitical and technological rivalries. According to well-informed sources and reports from international media, Moonshot AI is gearing up to acquire more GPU computing resources based on NVIDIA's Blackwell architecture, in anticipation of training its upcoming large-scale model, Kimi K4. Kimi K3, which currently holds the title of the world's largest open-source large-scale model in terms of parameter count, employs a Mixture of Experts (MoE) architecture and encompasses a total of 2.8 trillion parameters. It offers native support for multimodality and features a context window that can accommodate up to one million tokens. This model has demonstrated exceptional performance across various benchmark tests, setting new benchmarks for open-source large-scale models. Its open-source approach swiftly garnered enthusiastic feedback from the global developer community, amassing over 4,000 likes within a mere 30 minutes of its release and claiming the top spot on the trending list of the Hugging Face platform. Moonshot AI emphasized that full-stack open-sourcing not only reduces the barriers to accessing state-of-the-art intelligence and fosters industry innovation but also empowers users with greater control, privacy protection, and ownership over their data. With the unveiling of Kimi K3, Moonshot AI is expediting its path towards capitalization and intends to commence the final round of pre-IPO funding negotiations in August, with a pre-investment valuation target of $50 billion.