Chinese AI labs are diverging on whether to build native multimodal capabilities into their frontier models, according to a 36Kr report. Moonshot AI’s Kimi K3, Alibaba’s Qwen3.8-Max, and ByteDance’s Doubao-Seed-2.1 have committed to native multimodal training, while DeepSeek, Zhipu, and Tencent Hunyuan remain text-only on their latest general-purpose bases. Liang Wenfeng has said training AI well does not require a world model or even multimodality, but acknowledged that multimodality ultimately still needs to be done. Companies agree on the long-term value but disagree on timing and cost. The case for native multimodality centers on long agent tasks. For long-chain tasks, if you only receive code-level feedback, errors can accumulate and final outcomes become very poor. Vision is a more accurate feedback signal. As agents generate web pages, operate software, and check execution results, vision evolves from being an input the user hands the model to being the feedback the model uses to inspect its own work and adjust action. Kimi K3 reached the top of the Arena Frontend Code leaderboard at 1,679 points after release. Browser platform Puter inserted five visual deviations in test pages; Kimi K3 compared the target page with the running page screenshot and identified all five with no false positives. Moonshot AI calls this approach vision in the loop. External visual tools,