Moonshot AI Unveils Kimi K2.5 Open-Source Model

Chinese artificial intelligence company Moonshot AI unveiled its latest open-source large language model, Kimi K2.5, on Wednesday, positioning it as a major advancement in multimodal AI capabilities. The model, trained on approximately 15 trillion tokens encompassing text, images, and videos, promises superior performance in coding, reasoning, and agent orchestration compared to several proprietary rivals.

Moonshot AI, founded by former Google and Meta researcher Yang Zhilin, described Kimi K2.5 as building on its predecessor, Kimi K2Base, released in November. The new model employs a mixture-of-experts architecture with around 1 trillion parameters, distributed across multiple neural networks specialized for different tasks. This design activates only the relevant networks for a given prompt, reducing computational demands and hardware usage. Each expert network contains about 32 billion parameters, supported by a 400 million parameter vision encoder that converts multimodal inputs into processable embeddings.

Key enhancements include parallelized attention mechanisms, which speed up the model's review of relevant data, and a native multimodal framework that handles text,...