Weekly AI Roundup: Breakthrough Models, Open Source Innovations, and Global Reach (August 18-24)
This week’s AI highlights include Anthropic’s new Marshmallow and Melon models in early testing, Zhiyuan’s GLM-5.3 API launch, Qianwen Office’s...
The AI industry continues to surge forward with impactful developments this week, spanning cutting-edge models, open-source tools, enterprise integrations, and global expansion. Here’s a curated roundup of the most notable events from August 18 to 24.
Anthropic unveiled two new models, codenamed "Marshmallow" and "Melon," in its Early Access Program (EAP) for a limited group of developers. Community speculation suggests these models are not part of the Fable line but likely iterations of Opus/Sonnet or a new Haiku variant. Early feedback indicates Marshmallow outperforms Melon and even the current Opus 5 model in dialogue quality.
Zhiyuan launched the GLM-5.3 API, which scored 60 points in the AA Comprehensive Intelligence Index—matching closed-source flagships like Claude Fable5 and GPT-5.6 Sol. Specializing in complex coding, defensive cybersecurity, and long-haul tasks, GLM-5.3 offers lower costs and smaller parameters for wider accessibility. The API is now available, with model weights set to open-source this Friday.
OpenAI open-sourced the Codex execution framework Harness under Apache-2.0, providing CLI, SDK, and app-server components to embed AI agents into business systems. It also opened access to GPT-5.6 Sol’s 1 million-token context window (configurable via settings), though noting longer context may increase token usage and reduce ultra-long text precision.
Qianwen Office released MyContext, an open-source context infrastructure that transforms heterogeneous work data into agent-friendly profiles. This enables AI agents to understand user workflows, enhancing human-AI collaboration in decision-making and execution. Additionally, Qianwen Office integrated with Enterprise WeChat, completing support for China’s three major office platforms (DingTalk, Feishu, Enterprise WeChat).
MiniMax’s Design platform, built on its open-source multi-modal video model H3, turns creative intent into end-to-end video workflows (material generation, editing, subtitles). It supports commercial ads, educational content, and shorts, plus a 3D director’s desk for storyboarding and ComfyUI workflow integration.
Alibaba Cloud launched its third South Korean data center with Agentic AI services, expanding its global footprint (now 30 regions, 104 zones) to meet local AI/cloud demand. The company also leads the Asia-Pacific cloud market.
Anthropic’s Claude models autonomously designed proteins with a 26.8% success rate (vs industry average 10-15%), using open-source tools and taking 24h per target. Results are open-sourced on HuggingFace, though public Claude versions restrict this capability for safety.
The World Robot Congress in Beijing concluded with 44 new technologies (e.g., embodied world models) and 3000+ products. The Humanoid Robot Games saw "Lightning" set a 9.32s 100m record, and China’s industrial robot exports rose 13.2% YoY in Jan-Jul.
Sources
- Zhihu Column: Weekly AI Major Events (URL: https://zhuanlan.zhihu.com/p/2075293798876426832)