精选理由
Kimi K3和DeepSeek V4现在走了两条路：一个搞原生多模态，一个纯文本硬扛。长任务里能不能看懂画面，差距会越来越大。
Moonshot AI的Kimi K3、阿里Qwen3.8-Max和字节Doubao-Seed-2.1均采用原生多模态训练，而DeepSeek、智谱和腾讯混元仍坚持纯文本路线。在长程智能体任务中，视觉信息的实时参与成为决定性因素。这一差异将影响下一代模型在复杂环境中的表现。
原文 · pandaily
Between Kimi K3 and DeepSeek V4: Why Native Multimodal Capability Defines the Next Phase of Chinese Frontier Models
Moonshot AI Kimi K3, Alibaba Qwen3.8-Max, and ByteDance Doubao-Seed-2.1 commit to native multimodal training while DeepSeek, Zhipu, and Tencent Hunyuan stay text-only as vision-in-the-loop becomes decisive for long agent tasks.