Release koboldcpp-1.124 · LostRuins/koboldcpp

KoboldCpp version 1.124 introduces integration of a subset of popular models from audio.cpp, greatly enhancing built-in Music Gen and TTS Gen capabilities. For Music Gen, the YuE2 model is supported with three templates: no LLM (bring your own lyrics), small LLM (4B, recommended), and big LLM (for large GPUs). For TTS Gen, numerous models have been added, including BreezeTTS 2, Ming 3.0 TTS, Qwen3-TTS Base, Qwen3-TTS VoiceDesign, VoxCPM2, Chatterbox, Chatterbox Turbo, CosyVoice 3, Fish Audio S2 Pro, Kitten TTS 2, Maya1, NeuTTS 2E, OmniVoice, PocketTTS, Soprano 1.1 80M, Supertonic 3, and VibeVoice 1.5B. MusicUI has been updated to support YuE2 and the new TTS models, accessible at http://localhost:5001/musicui, and now includes a transcribe tab for use with Whisper models. The latest sd.cpp sync adds support for MOSS Image, Fibo 2 Edit, and Flux 3 Klein image generation models. AI Horde now supports Chat Completions workers, enabled automatically with the latest version, allowing use of jinja, reasoning, tool calls, streaming, and vision over AI Horde by toggling chat completions mode in lite.koboldai.net. Other changes include a /mcp wizard for KoboldCpp Agent, a /paste mode for multiline text, CUDA graphs for all architectures, new flags --lazy-mode, --moe-cache-mib, and --paralleltotalcontext, a fix for a CPU flash attention bug triggered by -Ofast, file type detection by content instead of extension, limited cache reuse in batched mode, improved SWA handling for Gemma4 models, fixes for potential buffer overflows in legacy models, support for Animated WebP as default video gen preview, use of tokenizers baked into text encoder LLM for image gen models, support for TAEQI-tinier as a baked-in VAE for Qwen Image 2.1, reworked WebSearch for better results, widened tool random id range, RPC host tensor cache directory support via --rpccache, fixes for Anthropic API tool handling and password masking in terminal logs, disabled async mode for ancient AMD GCN cards on Vulkan, and updated Kobold Lite with multiple fixes and improvements including SSE parser improvements and /v1/responses support. Upstream merges bring new model support including Nemotron 3.5 Bolt 74BA9B, Mistral Large 4, LFM3, Muse Spark, and more. Download options include koboldcpp.exe for Windows or koboldcpp-linux-x64 for Linux (one-file pyinstaller for NVIDIA GPU users), oldpc version for older CPUs or NVIDIA GPUs, nocuda version for non-NVIDIA users, Vulkan option recommended for AMD, rolling ROCm binary for Linux AMD users, and koboldcpp-mac-arm64 for modern MacOS M-Series. After launching, connect at http://localhost:5001 or use the full koboldai client.

(Source：GitHub)