Chinese content智谱发布GLM-5.3-FlashX，推理速度达200 tokens/s
智谱AI新出的GLM-5.3-FlashX模型，在国产加速器上跑得特别快，每秒能处理200个token，比之前版本效率高不少。
DailyVOL 2026.09.19308 items
智谱GLM-5.3-FlashX推理速度达200 tokens/s，AI代理与多模态技术持续突破
Singapore·Saturday, 19 September 2026·DAILY · 08:00 SGT
2 Sep 2026. 智谱AI推出的GLM-5.3-FlashX模型，在约10万颗国产加速器上实现每秒200个token的推理速度。该模型基于320亿参数的混合专家（MoE）架构，并使用基础设施代理优化了服务堆栈。与此同时，ActObs方法通过监督动作和观测令牌，在Qwen3-4B上提高了终端基准测试的通过率。FAMOS模型则通过从稀疏点云预测可移动部分和关节参数，实现了3D物体建模的新突破。此外，华为发布的世界首个NPO超节点‘昇腾960 Super Node’预计将于2027年第三季度推出，提供7.2T容量。亚马逊Bedrock的AgentCore运行时新版本通过内存回收和优化，实现了更一致的冷启动速度。北京发布‘词元经济十条’，推动词元工厂建设，并支持推理专用芯片、模型轻量化等技术研发。Micron表示内存供应 meaningful until 2028，当前短缺影响所有市场。甲骨文花费数十亿美元帮助其他企业运行AI，但自家公司内部推广AI仍面临挑战。Open-source设计工程技能被分享，用于提升编码代理的UI细节处理能力。
01
Models
3 itemsActObs changes how agents explore under RL
This research provides a new method for improving the exploration capabilities of AI agents, which could lead to more robust and effective systems in various applications.
FAMOS: Feed-Forward 3D Articulation Modeling from Sparse Observations
This research introduces a new method for 3D object modeling from sparse data, which could be useful for applications like robotics and augmented reality.
02
Products
3 itemsHuawei releases world's first NPO-based super node '昇腾 960 Super Node', launching Q3 2027
For users needing high-performance AI infrastructure, the Ascend 960 Super Node offers a significant reduction in power consumption and improved system reliability compared to traditional setups.
IT之家Source
Chinese contentAmazon Bedrock 新版 AgentCore 运行时发布
亚马逊发布了 Bedrock 的 AgentCore 运行时新版本，能更高效地处理生产环境中的 Agent，比旧版本更快更稳定。
Claude Code Projects update: Claude Tag architecture + Slack Thread feature
This update makes Claude Code more efficient for managing complex coding projects by automating task delegation and parallel processing, which can save time for developers.
03
Industry
3 itemsMicron says meaningful memory supply won't ramp until 2028
For those in the tech industry, this explains why memory prices remain high and why AI model development is facing bottlenecks, as the supply chain is not keeping up with demand.
IT之家Source
Chinese content甲骨文烧数十亿美元帮其他企业运行 AI，公司内部推广 AI 却并不顺利
甲骨文自己用 AI 写代码快了，但公司内部推广 AI 还不顺利，成本和流程问题都出来了。
IT之家Source
04
Papers
3 itemsFast Cross-Strength Multi-Contrast Brain MRI Translation
This research provides a faster and more efficient method for translating brain MRI images, which is crucial for medical analysis and research.
QoS-Aware Federated Learning for Smart Vehicles
This research is relevant for anyone working on AI systems in vehicles, as it provides a framework to optimize federated learning under real-world network constraints.
Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision
For those working on robotics or AI, this research offers a practical solution for creating more efficient and effective robotic memory systems.
05
How-to
3 itemsChinese contentCua 发布 jev-use，将「快速电脑使用」拆分为两层架构
Cua 发布了 jev-use，把「快速电脑使用」拆成两层架构，Jev 负责决策，Cua Driver 负责执行，这样效率高很多。
Chinese content亚马逊 SageMaker AI 支持部署 Hugging Face 模型
亚马逊 SageMaker AI 现在支持部署 Hugging Face 模型了，指向模型就能自动创建实时端点，包含服务容器、自动扩展和 CloudWatch 警报。
Open-source design engineering skills for coding agents
This is a practical guide for developers on how to create and maintain a structured set of design rules for coding agents, which can improve the quality and consistency of UI development.
308Stories
145First-party
93New models
79Sources
AITOP · generated by the editorial system