Release koboldcpp-1.123 · LostRuins/koboldcpp

GitHub
KoboldCpp 1.123 launches the koboldcpp.net community site, adds Qwen Image 2.1 support, major Agent and batched mode improvements.

Summary

KoboldCpp 1.123 introduces koboldcpp.net, an official community website created with @Autumnlight02 to help newcomers and counter fake sites, while official downloads remain on GitHub. It adds support for the Qwen Image 2.1 model for image generation and editing via a ready-to-use .kcppt template, thanks to @wbruna. The integrated KoboldCpp Agent receives major updates: new /confirm approval modes (auto, on, off) configurable globally or per tool, /save and /load commands with --agentload for sessions, /maxsteps, /usage, /autocompact, and /paste commands, Esc to interrupt prompts, fixes for system prompt location and stdin blocking, MCP tool calls limited to localhost, and a dedicated Agent tab. Batched mode now supports all standard samplers, grammar, tool calls and vision, with --parallelrequests for instantaneous responses. Other fixes include rep_pen_range OOM, grammar parsing, vision for oldpc, idle unload, and new Spark X2.5 template; the --jinjatools flag is deprecated in favor of automatic Jinja tools via --jinja, and --multiuser is replaced by --queuelimit. Miscellaneous cleanup, Termux launcher improvements, remotetunnel path fix, shared timeout, updated docs, and Kobold Lite changes are included, plus multiple custom OpenAI endpoints can be saved as favorites. Users can download koboldcpp.exe for Windows or koboldcpp-linux-x64 for Linux, with oldpc, nocuda, Vulkan, ROCm, and MacOS binaries available depending on hardware, and run via command line or GUI.

(Source:GitHub)