Local LLM runtime
KoboldCpp
KoboldCpp is a AGPL-3.0-licensed local-LLM desktop app for Linux · macOS · Windows, running on CUDA · CPU · Vulkan · ROCm, with an OpenAI-compatible API. Ease of use 4/5 in our sourced ratings.
- License
- AGPL-3.0
- Operating systems
- Linux · macOS · Windows
- Backends
- CUDA · CPU · Vulkan · ROCm
- Install
- binary / source
- Ease of use
- 4/5
- Latest version
- v1.116.1
- GitHub stars
- 10,973
- Engine
- llama.cpp
- Graphical app (GUI)
- Server mode
- OpenAI-compatible API
- CPU offload
- Multi-GPU
- Speculative decoding
- LoRA support
- KV-cache quantization
As of 2026-07-08 · source · GitHub
Quant-format support
What the feasibility gate checks before any VRAM math — per format and backend.
| Format | Backend | Supported |
|---|---|---|
| gguf | CPU | yes |
| gguf | CUDA | yes |
| gguf | Metal | yes |
| gguf | Vulkan | yes |
Software that runs on KoboldCpp
Frontends, assistants and harnesses that can use KoboldCpp as their local backend.
Will your machine run a model with KoboldCpp?
Check with KoboldCpp preselected →Support facts come from each runtime's own documentation, stamped and dated — how the math works.