China AI Native Industry Insights – 20260812 – Alibaba | Tencent | Zhipu AI | more
Today’s digest spotlights developments from Alibaba and Tencent, with the release of Qwen-MM-Plugins and WorldClaw making headlines. These advancements highlight the growing significance of multimodal-native AI agents and agentic frameworks for open-world generation. The Qwen-MM-Plugins are designed to enhance multimodal AI capabilities, while WorldClaw provides a framework for converting text to immersive 3D environments. Concurrently, ZCode sees a significant upgrade, solidifying its position as GLM’s top coding tool with enhanced features for autonomous task delivery. This marks a pivotal step in integrating AI with dynamic world-building and task automation. Discover more in Today’s China AI Native Industry Insights.
1. Alibaba Qwen Releases Qwen-MM-Plugins to Enable Multimodal-Native AI Agents
Alibaba’s Qwen team has released Qwen-MM-Plugins, a plugin framework designed to extend existing agent harnesses with multimodal-native capabilities. The plugins allow AI agents to natively read and process images, videos, and documents, edit videos, and work with 3D and CAD content. The release bridges the gap between multimodal language models and fully multimodal agentic workflows. Qwen-MM-Plugins is aimed at developers building AI agents who want to incorporate rich media understanding and manipulation beyond text.
Read more: https://github.com/QwenLM/Qwen-MM-Plugins
Video Credit: @Alibaba_Qwen on X
2. Tencent Hunyuan Releases WorldClaw, an Agentic Framework for Text-to-3D Open-World Generation
Tencent Hunyuan has introduced WorldClaw, a fully agentic, coarse-to-fine framework that generates large-scale, explorable 3D open worlds from open-ended text prompts. Planning agents translate a text prompt into a structured specification of regions, terrain, assets, materials, and spatial relations, then build a globally coherent scene composed of independent textured meshes. Unlike video or Gaussian Splatting approaches, every scene produced by WorldClaw consists of editable, game-ready 3D assets with high-quality geometry and textures suitable for free-viewpoint exploration. The system leverages Claude Opus 4.8 as the underlying agent model and integrates foundation models including GPT-Image-2, SAM3, SAM3D, and Hunyuan3D, running on NVIDIA H20 GPUs with Blender 5.1.1 for scene assembly. WorldClaw addresses the core challenge of jointly maintaining global spatial coherence and rich local content in large-scale 3D scene generation.
Read more: https://arxiv.org/abs/2608.05248
Video Credit: @TencentHunyuan on X
3. ZCode Gets Major Upgrade as GLM’s Top Coding Harness, Enabling Autonomous Task Delivery
Zhipu AI’s ZCode, a coding harness optimized for its GLM models, has rolled out a major upgrade adding four new capabilities: Goal, Subagents, Remote Control, and idle-time tasks. With Goal mode, developers simply set a clear, verifiable objective (e.g., “keep first-screen load time under 2 seconds and pass all existing tests”), and ZCode autonomously breaks down the task, edits code, runs commands and tests, and checks whether the goal has been met — continuing through additional rounds until it succeeds, without requiring constant human prompting. Paired with Subagents, a single agent can spin up parallel sub-agents — a General-purpose type for code edits and command execution, and an Explore type for read-only investigation — turning one agent into a coordinated team, which is especially useful for full-stack feature work, cross-file refactoring, and complex bug hunts. Remote Control lets developers monitor and control desktop tasks from their phone (including via WeChat, Feishu, and Lark), while code execution stays on the original local, SSH, WSL, or Docker environment. According to Zhipu’s internal
Read more: https://mp.weixin.qq.com/s/NwVPQ8zsyIp8BC0Z1dN90g
Video Credit: The original article
That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.