Today's AI trends
llama.cpp’s b10829 release corrected Qwen3-Next gated delta-net normalization to match the reference epsilon placement; it also illustrates why a runnable converted model still needs cross-runtime validation. Teams upgrading inference engines should pair performance testing with architecture-specific numerical and behavioral regression checks. Release b10829
Claude Code v2.1.261 added configurable limits for command and task output before saving it to a file, file-based subagent system prompts, organization-policy diagnostics, and a skill doctor that reports unused skills and their context cost. The practical implication is to treat instructions and tool output as managed Context resources rather than free text in long-lived agent sessions. Claude Code v2.1.261 release