This week's technical updates feature significant shifts in model availability and protocol architecture. Moonshot AI released the Kimi K3 weights on HuggingFace, a sparse Mixture-of-Experts (MoE) model utilizing MXFP4 quantization with 2.8 trillion total parameters and a 1M token context window. While it ranks high on the Frontend Code Arena, early evaluations noted a 51% hallucination rate on certain agentic tasks, and its scale necessitates H100-level infrastructure for production. In contrast, Z.ai's GLM-5.2 offers a more cost-effective 753-billion parameter alternative under an MIT license, scoring 81% on Terminal-Bench 2.0 compared to K3's 88.3%. The Model Context Protocol (MCP) underwent a major revision on July 28, transitioning to a stateless architecture. This allows servers to operate behind round-robin load balancers by routing via Mcp-Method headers rather than maintaining sticky sessions. New features include multi-round-trip requests for asynchronous patterns and TTL-based caching for tools. Additionally, OfficeCLI launched as an Apache 2.0 C# tool, providing AI agents with a single-binary interface to read, write, and render .docx, .xlsx, and .pptx files without requiring Microsoft Office. It supports 350+ Excel functions and uses a resident mode with named pipes to maintain documents in memory, optimizing agentic loops. Finally, VS Code's July 2026 update introduced multi-chat sessions and vision support for PDFs and images within Copilot Chat.