koboldcpp
koboldcpp, single-binary local LLM runner built on llama.cpp with a built-in chat UI.
Release history
koboldcpp 1.123.1
KoboldCpp 1.123.1 adds Qwen Image 2.1 support, major agent improvements with granular approval modes and session persistence, full sampler/grammar/vision support in batched mode, and fixes macOS GPU usage regression
koboldcpp 1.123
KoboldCpp 1.123 adds Qwen Image 2.1 support, major agent improvements with approval modes and session management, and batched mode now supports samplers, grammar, tool calls, and vision
koboldcpp 1.122.1
KoboldCpp adds an integrated agent with 9 built-in tools and 2k token system prompt for autonomous task execution across compatible backends
koboldcpp 1.120
Adds DirectIO model load mode; Fixes assistant gen prefill trigger and failsafe mode bug. Mlock and mmap can now be combined
koboldcpp 1.119
Video generation and I2V support with Minimax H3
koboldcpp 1.118.1
Row Split option removed; Selecting it now defaults to tensor split
koboldcpp 1.117.1
KoboldCpp 1.117.1 fixes thinking trace display, improves multimodal image placement, adds full Ollama compatibility with embeddings and tool calling, and reworks group chat with multi-character avatars
koboldcpp 1.117
Fixes terminal output for thinking traces; Multimodal images now inserted inline instead of at context top
koboldcpp 1.116.1
Koboldcpp 1.116.1 adds Krea 2 Turbo, Ideogram 4, and Boogu Edit image generation models, MCP server support in the web UI, and fixes MTP performance issues
koboldcpp 1.116
Adds Krea 2 Turbo support
koboldcpp 1.115.1
Koboldcpp 1.115.1 adds Anthropic API support with multimodal and tool calling, Gemma4 UV models, MTP speculative decoding, video generation improvements, and sampling speedups
koboldcpp 1.115
Implements full Anthropic /v1/messages API support including multi-turn conversations
koboldcpp 1.114.1
Koboldcpp 1.114.1 adds experimental parallel text generation, multi-GPU RPC support, LTX2.3 video generation, and support for Microsoft Lens, HiDream o1, LongCat, and Ernie image models
koboldcpp 1.114
Experimental parallel text generation, RPC backend for distributed GPU sharing across networks, LTX2.3 video generation support, and multiple new image generation models added
koboldcpp 1.113.1
Breaking change: split mode now set independently to accommodate experimental tensor split. Existing launch scripts may need adjustment
koboldcpp 1.113.2
KoboldCpp 1.113.2 fixes router mode regressions, splits tensor configuration independently, adds LoRA directory selection, and improves VAE decoding speed