Latest in AI
KoboldCpp 1.124 Adds YuE2 Music and More Local Voice Models
The update puts more audio models inside one local interface, while hardware and model choice still determine what you can run.
KoboldCpp released version 1.124 on October 31 with YuE2 music generation and more text-to-speech models. For people running AI locally, the useful change is having more of the creative workflow available in the same application.
What changed in KoboldCpp 1.124
The release notes describe an integration of selected models from audio.cpp. Voice options include BreezeTTS 2, Ming 3.0 TTS, Qwen3-TTS and VoxCPM2. Support for voice cloning or written instructions varies by model.
YuE2 has three templates: music generation with your own lyrics, a version paired with a small 4B language model for lyrics, and a larger language-model option intended for large GPUs. MusicUI also gains a Whisper transcription tab, useful when a voice model needs a transcript of reference audio.
KoboldCpp 1.124 is an application update, not the original launch of these models.
How the local setup works
KoboldCpp's project documentation describes a free, open-source application with a bundled browser interface and local APIs. It offers ready-to-run binaries for Windows, Linux and Apple Silicon Macs. Model files are downloaded separately, and their size and context requirements determine memory needs.
The audio integration builds on audio.cpp, a C++ inference engine powered by ggml. That project supports several audio workloads without a Python dependency. KoboldCpp integrates a subset, so audio.cpp supporting a model does not automatically mean KoboldCpp supports it too.
For the new music and voice features, the release directs users to MusicUI at http://localhost:5001/musicui after launching the application. Start with the template that fits the task and available hardware, rather than assuming the largest lyrics model improves every music generation.
Fewer interfaces, the same evaluation work
A local creative workflow can become awkward when every model needs its own environment and interface. Bringing speech and music into the same application can reduce that setup work. It also makes it easier to compare a few models before deciding which belongs in a larger workflow.
But integration does not establish output quality, low latency or suitability for a particular GPU. The release provides no common benchmark that makes its supported voice models directly comparable. Voice-cloning support also does not grant permission to copy someone's voice; use reference recordings you are authorised to use.
For developers building on local AI, I would watch how easily users can move from a working sample to a repeatable task. Adding another supported model is useful. Making the right model straightforward to load, evaluate and keep using is what makes the application worth returning to.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain model integrations and product launches through clear technical stories that developers can understand and share.