Official account of Tencent Technology. A platform for publishing and analyzing Tencent's technological innovations and cutting-edge developments.
The article argues that in the AI era, writing specifications, rules, and evaluation sets forces engineers to externalize tacit knowledge, turning AI from a tool into a mirror that reveals hidden assumptions, and warns that this legibility brings both powerful benefits and profound risks.
The author evaluates Marvis, an AI‑powered desktop assistant that bundles six specialized agents—fast terminal scheduling, autonomous planning, cross‑modal task chains, a visual agent workspace, vibecoding for code, and desktop organization—showcasing rapid local execution, privacy‑preserving design, multi‑agent coordination, and future mobile integration.
Analyzing 3,632 high‑severity GitHub Advisory reports from 2025‑2026, the authors reveal a sharp rise in business‑logic flaws—especially in high‑star projects—prompting a redesign of vulnerability‑detection benchmarks, and introduce VulnGym, a real‑project, white‑box dataset with 400+ paths and detailed entry‑point, trace, and critical‑operation annotations.
This article walks through a two‑month source‑code study of vLLM, explaining how token‑level scheduling, continuous batching, and the Paged Attention mechanism reshape tensor dimensions to turn large‑model inference into a compute‑bound, high‑throughput process while managing GPU memory efficiently.
The Tsinghua‑Tencent Mixed‑X team captured the MLSys 2026 MoE inference optimization championship by analyzing NPU bottlenecks, redesigning data movement, applying expert‑level sharding, continuous DMA, PSUM batching, and an Agent‑based optimizer, achieving a 4.1× end‑to‑end speedup while preserving bit‑level output fidelity.
On April 23, Tencent released the open‑source Hy3 preview, a 295 B‑parameter hybrid expert model with 21 B active parameters and 256K context length, delivering substantial gains in complex reasoning, instruction following, code and agent tasks, achieving 40 % faster inference, lower costs, and strong benchmark results across Tencent’s AI products.
When AI coding hype focuses on front‑end page generation, the real challenge is whether AI can be trusted to write code for a million‑line, high‑availability CDN backend; this article details Tencent’s systematic exploration, a 20‑day Rust proxy prototype, a five‑layer Harness Engineering framework, and concrete data showing both breakthroughs and remaining risks.
The article outlines a multi‑stage AI‑driven automation roadmap that extends from requirement gathering through technical solution generation, code creation, testing, and deployment, detailing challenges such as workflow standardization, knowledge‑base construction, skill reuse, and quality assurance, and presents concrete metrics showing up to 80% efficiency gains.
This article walks through a complete Agentic Engineering workflow for a backend Go service, showing how AI‑driven Skills and Commands automate requirement capture, clarification, planning, parallel development, code review, deployment, log analysis, MR creation, AI‑assisted review, automated fixes, and final merge, while highlighting token consumption and practical lessons.
This article translates and consolidates three LWN analyses of the Linux swap subsystem modernization led by TencentOS kernel engineer Kairui Song, detailing the introduction of swap tables, removal of the swap map, virtual swap concepts, code changes, performance improvements of up to 20 % and the broader impact on the kernel community.
