Weekly Tech Roundup (May 18‑24): Does Tencent’s Marvis Bring Six AI Assistants to Your Desktop?

Tencent’s App Store team is piloting an OS‑level AI assistant called Marvis, which embeds directly into the operating system with Explorer‑like permissions, can understand file types, auto‑categorise images via OCR, modify system settings, diagnose hardware, and run offline via a local‑plus‑cloud routing model; a privacy mode runs entirely on‑device (source: 腾讯科技).

Marvis features a cartoon “Agent Studio” with a main agent wearing a red scarf and several specialized sub‑agents, focusing on document parsing, image classification, cross‑app actions, and system troubleshooting. Compared with open‑source frameworks like OpenClaw and Hermes, Marvis emphasizes office‑document handling, claiming to surpass 95% of market products (source: 腾讯科技).

In logistics, a robot deployed in a SF Express postal warehouse earned the nickname “first‑place in the national college entrance exam” after demonstrating superior performance in real‑world sorting tasks (source: article).

Chinese researchers from Zhipu, Yuxun, and Tsinghua introduced the ZCube networking architecture in the GLM‑5.1 production cluster, achieving a 15% increase in inference throughput and a 40.6% reduction in P99 token latency without adding GPUs or changing servers, while cutting switch and optical module costs by about one‑third (source: 量子位). ZCube replaces traditional Fat‑Tree/Clos topologies with a flat two‑hop design, reducing congestion probability; experiments show the same hardware and software stack gains over 15% throughput when only the network is re‑engineered (source: 量子位).

OpenAI announced that a general‑purpose inference model independently disproved the 80‑year‑old “unit distance” conjecture in discrete geometry, delivering a new point‑set construction with a 0.014 exponent improvement, and demonstrated cross‑domain reasoning using algebraic number theory (source: 新智元).

Google I/O 2026 unveiled a suite of agents: Gemini 3.5 Flash (four‑times faster than the prior Pro at half the price, built for long‑running tasks), Gemini Spark (a 24/7 cloud personal agent that extracts information from Gmail and Docs), Antigravity 2.0 (a multi‑agent orchestration platform challenging Claude Code and Codex), and CodeMender (an agent that auto‑fixes code vulnerabilities). Gemini app was redesigned for visual answers, Daily Brief generates personalized briefings, Gemini Omni integrates Nano Banana, Veo, and Genie for multimodal generation with SynthID watermarks, and the search box merges AI Mode and AI Overviews (source: APPSO).

AMD’s Shanghai AI conference highlighted that scaling GPU count alone no longer solves cost and complexity; the company promotes a full‑stack “GPU + CPU” approach, open‑source ROCm ecosystem, and a China‑focused AI developer program to aid model adaptation on AMD hardware (source: 机器之心).

WeChat Read launched an official Skill that exposes a user’s reading history—bookshelves, progress, highlights, notes—to agents via an API, enabling queries like “What books have I been reading lately?” and generating personal knowledge bases; the Skill is read‑only but showcases a new “user‑data AI” paradigm (source: 硅星人Pro).

GitHub’s top‑starred AI projects are now being used in production, and Feishu’s CLI, after just two months, amassed over 10 000 stars, offering a three‑layer stack for agents: high‑level shortcuts, structured API commands covering 17 business domains, and low‑level raw API access to 2 500 endpoints, positioning Feishu as an AI‑native operating system for organizations (source: CSDN).

At Alibaba Cloud 2026, the Qwen 3.7‑Max model topped the Arena global benchmark, outperforming domestic rivals and approaching GPT‑level performance. Optimised for agent tasks, it achieved SOTA on SWE‑Pro and Terminal Bench 2.0‑Terminus, and demonstrated autonomous code generation, tool usage, and long‑run task execution, including a 35‑hour self‑optimisation of an attention kernel on a custom AI chip (source: 机器之心).

Overall, the week highlighted a shift toward OS‑level AI assistants, network‑centric performance gains, multimodal agent ecosystems, and the growing importance of open‑source tooling and data‑centric AI strategies across the industry.