Moonshot AI releases Kimi 2.5 model that claims to outperform LLMs from OpenAI, Google

Kimi 2.5

• Moonshot unveils Kimi K2.5 AI with enhanced coding and multimodal features.
• Kimi K2.5 boasts better video generation and agentic features than US rivals.
• Developers can access Kimi K2.5 via open-source IDE, Kimi.com, app, and API

Did our AI summary help?

Chinese companies are accelerating the rollout of new artificial intelligence (AI) models as competition against US-based rivals such as OpenAI, Anthropic and Google escalates. While the US companies are pushing out these models with increasing budget, like the latest Gemini and Claude, the Chinese tech brands like Moonshot and DeepSeek are instead relying on putting more refinements at a much lower cost.

Along similar lines, Moonshot has released Kimi K2.5, the first AI LLM model from the company that combines strong reasoning, coding, and multimodal abilities to provide a more enriching experience than its American counterparts. It claims to have video-generation and agentic capabilities that outperform the likes of Gemini, ChatGPT, or Claude, and also understands text, image, and video uploads.

Moonshot states that the model was trained on 15 trillion mixed visual and text tokens, and it’s good at coding tasks and handling agent swarms. Further, the model supports both visual and text inputs, letting users generate a draft of a new website based on a recorded video of an existing website or feed it images or videos and ask it to make a similar interface shown in those media files. Although Moonshot did not publicly disclose K2.5’s parameter count, its predecessor, the Kimi K2 model, had 1 trillion total parameters and 32 billion activated parameters, with the latest model expected to surpass that.

Kimi K2.5's coding capabilities have been made available through an open-source platform called Kimi Code, which is an integrated development environment (IDE) like Cursor, VSCode, and Zed. It is also available through Kimi.com, the Kimi App, and the Kimi API. The company highlights that developers can run the model locally, fine-tune it on private datasets, or even operate it through Moonshot’s API at a much lower cost compared to its US-based rivals.