Release koboldcpp-1.122 · LostRuins/koboldcpp

GitHub
KoboldCpp 1.122 introduces an integrated AI agent harness, batch controls, audio generation, and numerous bug fixes and improvements.

Summary

The release of koboldcpp-1.122 brings a headline feature: an integrated KoboldCpp Agent, a lightweight agentic harness with 9 built-in tools serving as a replacement for tools like Opencode, Codex, or Claude Code. It uses a compact 2k-token system prompt, supports third-party backends and OpenAI-compatible endpoints, integrates MCP tools, and includes three approval modes for tool calling. A minimum of 28k context, 8k generation, and 12GB VRAM is recommended. Other key changes include independent ubatchsize control from batchsize (with pipeline parallel auto-enabling), increased default context size, a new --autoswapthreshold flag to control memory swapping, tool calling parser fixes, and a syncing of upstream sd.cpp. The release also adds audio code generation for lyrics (two-step process), improved music UI, smarter max output token handling near context limits, updated Kobold Lite, and various streaming and UI improvements. Note for package maintainers: koboldcpp.sh now optimizes for the target platform, so KCPP_PORTABLE=1 must be set for cross-platform builds. Download options cover Windows (NVIDIA/CPU), Linux (CUDA/Vulkan/ROCm), macOS (M-Series), and legacy hardware versions.

(Source:GitHub)