The Kimi K2 artificial intelligence model, launched by Moonshot AI, a company backed by Alibaba Group, is rapidly gaining traction and receiving accolades from industry experts. Downloads of Kimi K2 have surged, reaching 145,000 on Monday, up from 76,000 the previous Friday, according to data from the AI and machine-learning platform Hugging Face. This large language model, developed by the Beijing-based start-up, utilizes a mixture-of-experts architecture, featuring 1 trillion parameters with 32 billion activated per inference. In comparison, another model, DeepSeek-V3, has 671 billion parameters.
Henning Steier, the chief marketing officer of Bluespace Ventures, highlighted in a LinkedIn post that while companies like OpenAI invest heavily in computational resources, Moonshot’s Kimi K2 offers a more cost-effective approach to training and inference, potentially signaling a shift in AI development strategies. Unlike OpenAI’s GPT and Anthropic’s Claude, which require monthly subscriptions, Kimi K2 is available for free through its app and browser interface.
The mixture-of-experts approach in machine learning divides the AI model into separate sub-networks, each focusing on a subset of the input data, which significantly reduces computation costs during training and enhances performance during inference. Moonshot claims that the development of Kimi K2 was achieved at a fraction of the cost typically incurred by larger AI companies.
