The time difference between Kimi K3 and DeepSeek V4 is the native multi-mode

Text | Li Jiefeng Editor | Zhang Yuxin "If long-chain tasks only receive feedback at the code level, errors may continue to accumulate, and the final effect will be very poor." Talking about the significance of native multimodality, a multimodal researcher said, "Visual is a more accurate feedback and is closer to the user's intention." In the past year, Coding and Agent capabilities have continuously rewritten the rankings of large models, and have become one of the scenarios where AI can realize commercial value the fastest. At the same time, as Agents begin to take over more long-chain tasks, more and more general large models begin to match native multi-modal capabilities. In January this year, an enterprise usage report released by OpenAI showed that among the three most commonly used ChatGPT tools in R&D positions, image uploading ranked third after search and data analysis. As Agents begin to generate web pages, operate software, and check running results, the role of vision is gradually evolving: it is no longer just a picture handed to the model by the user, but also begins to become feedback for the model to check the work, find errors, and adjust actions. This past July, Dark Side of the Moon released Kimi K3. This MoE (Mixed Expert Model) with a total parameter of 2.8 trillion and supports 1 million token contexts. In addition to its coding and long-range agent capabilities, its excellent native multi-modal capabilities are another label of K3. The so-called native multi-modality allows images, text and other data to jointly shape the main model from the pre-training or continuous pre-training stage, and continues to be optimized in post-training, and finally participates in the perception and decision-making of the Agent (agent). After K3 was released, it topped the Arena Frontend Code list with 1679 points. The list is ranked by interactive web pages generated by the user's blind selection model. It is judged not only whether the code can run, but also the final effect of the page. The browser development platform Puter once created 5 visual deviations in a test webpage. K3 compared the screenshots of the target page and the running page and found all of them without false alarms. Thanks to its excellent native multi-modal capabilities, K3 can understand the visual problems after the code is run, providing a basis for the next round of modifications. Ki