Our daily AI roundup brings together key technical developments across artificial intelligence, computing hardware, and agent tooling. Research teams and infrastructure providers continue to publish performance benchmarks, while developer platforms release new protocols for autonomous systems.

Hardware and Physical World Infrastructure

Perplexity published research on its Portable Computer architecture, which uses an on-device 27B model for local execution. The base model achieved 82.6% on real knowledge work tasks, whereas the post-trained PPLX 27B variant scored 85.4%. In enterprise hardware, Cerebras introduced its CS-4 accelerator on the Nexus rack-scale platform, while previewing the CS-5 architecture aimed at higher token-generation speeds. Meanwhile, robotics data platform Index emerged from stealth, reporting 264,000 downloads across 100 countries and 30 minutes of physical data uploaded per second.

Corporate Deployments and Enterprise Tools

Skild AI initiated deployments of its S1 model with a select group of industrial partners before opening broader access over coming months. In workflow automation, Factory updated Droid with Connectors and Personas to integrate with Databricks, Google, Linear, and 50 other tools. Furthermore, OpenAI added website login capabilities inside ChatGPT Work across web and mobile interfaces for Plus, Pro, and Business tiers, streamlining navigation inside external apps.

Model Benchmarks and Compute Clustering

Alibaba’s Qwen3.8-27B secured first place among open models on Arena.ai’s Image-to-WebDev benchmark with a score of 1,574, priced at $0.40 per million input tokens and $3 per million output tokens. In server reliability, NVIDIA introduced a Dynamo shadow engine preview that restored GLM-5.2 engine capacity in 7.3 seconds, nearly 39 times faster than cold restarts. Additionally, EXO Labs collaborated with Apple to cluster computers via Thunderbolt 5 RDMA, enabling high-speed inference for Kimi K3 and GLM-5.3 models.

Autonomous Agents and Web Protocols in the Daily AI Roundup

Nous Research expanded the Hermes Agent catalog by integrating 44 new connections configurable within its desktop interface. At the same time, OpenAI Developers introduced WebMCP, an experimental open standard designed to let web applications expose tools directly to autonomous agents. Developers can participate in the WebMCP building challenge before the September 3 deadline, with selected projects announced on September 23. This daily AI roundup highlights how interconnected models and hardware architectures continue to evolve.