Key Takeaways
- AGPL 3.0 license β free to use, copyleft if you modify and redistribute it as a network service
- Built on llama.cpp (MIT-licensed) underneath, with its own bundled UI and extra features added on top
- Single executable per platform: Windows, Linux x64, macOS ARM64 β no installer
- KoboldAI Lite, the bundled web interface, is built for story writing and roleplay: editable output, memory/world-info fields, and multiple chat/adventure modes
- GPU acceleration via CUDA (NVIDIA) or Vulkan (cross-vendor, recommended path for AMD); a rolling ROCm build exists for Linux; CPU-only also fully supported
- Also bundles Stable Diffusion image generation and Whisper speech-to-text in the same executable
π In One Sentence
KoboldCpp is a free, AGPL 3.0-licensed, single-file local LLM application built on top of llama.cpp that runs GGUF models with zero installation and ships with the KoboldAI Lite web interface for story writing and roleplay.
π¬ In Plain Terms
Download one file, double-click it, and a browser tab opens with a chat and story-editing interface already built in β no separate installer, no command-line setup, no wrapper app needed on top of it.
πNote: The only official download source is the GitHub Releases page. Files with the same name hosted elsewhere are not verifiable as the genuine build.
What Is KoboldCpp?
KoboldCpp is a free, single-file program that runs GGUF-format language models locally, built on top of the llama.cpp inference engine and maintained by a pseudonymous developer known as LostRuins. It descends from the original KoboldAI project's interface conventions, adapted to run on llama.cpp's backend instead of a Python-based inference stack.
The project is hosted at github.com/LostRuins/koboldcpp and released under the GNU Affero General Public License version 3 (AGPL 3.0) β a copyleft license, distinct from the permissive MIT license llama.cpp itself uses underneath. Releases ship frequently; the version current as of testing (late August 2026) was v1.120.
- Distributed as one executable per platform β koboldcpp.exe (Windows), koboldcpp-linux-x64 (Linux), koboldcpp-mac-arm64 (macOS Apple Silicon)
- No installer, no Python environment, and no Docker required to run it
- Wraps llama.cpp for the actual model inference, then adds its own server, web UI, and API layer on top
- Ships a nocuda build (CUDA-free, Vulkan-capable) and an oldpc build (CUDA11 + AVX1) for older hardware, alongside the main CUDA-enabled build
What Is KoboldAI Lite, and What Does It Add?
KoboldAI Lite is the web interface bundled directly inside the KoboldCpp executable β it opens automatically in a browser tab once the server starts, with no separate install step. It is built specifically around long-form writing and roleplay, not just turn-by-turn chat.
- Editable output: generated text can be edited in place, not just regenerated, which matters for collaborative story writing
- Memory and World Info fields: persistent context blocks the model always sees, used to keep character details or story facts consistent across a long session
- Multiple modes: standard chat, a "Story" mode for collaborative fiction, and an "Adventure" mode styled after text-adventure games
- Author's Note and instruct-formatting controls for steering tone and style without editing the underlying prompt template by hand
- Runs entirely from the same executable and the same local server β no separate frontend app or browser extension needed
How Do You Download and Run KoboldCpp?
Download the correct file for your platform from the official GitHub Releases page, then run it directly β there is no install step. Picking the right build for your hardware is the only real decision involved.
- 1Go to the KoboldCpp Releases page on GitHub β this is the only official download source.
- 2Pick your platform build: koboldcpp.exe for Windows, koboldcpp-linux-x64 for Linux, or koboldcpp-mac-arm64 for macOS on Apple Silicon.
- 3Pick your GPU path: the main build includes NVIDIA CUDA support; AMD users are pointed by the project to try the Vulkan option in the nocuda build first; a rolling ROCm build exists separately for Linux; older hardware without modern instruction sets should use the oldpc build (CUDA11 + AVX1).
- 4Download a GGUF model file (from Hugging Face) if you do not already have one.
- 5Run the executable, select the GGUF file and a GPU-layer offload amount in the launcher window, and start the server.
- 6A browser tab opens automatically to the KoboldAI Lite interface, already connected to the running local server.
Does KoboldCpp require installation?
No. It is a single executable file β download it and run it directly. There is no installer, no separate Python environment, and no Docker container involved.
Can KoboldCpp run without a GPU?
Yes. CPU-only operation is fully supported; a GPU (via CUDA, Vulkan, or ROCm) only accelerates generation speed, it is never required.
Which GPUs and Hardware Does KoboldCpp Support?
KoboldCpp supports NVIDIA, AMD, and CPU-only setups through separate build variants, rather than one universal binary auto-detecting everything. Picking the matching build for your hardware is required to get GPU acceleration working.
- NVIDIA β CUDA acceleration is built into the main release binaries
- AMD β the project's own guidance is to try the Vulkan backend in the nocuda build first, since it has the broadest compatibility; a separate rolling ROCm binary is also maintained for Linux
- Cross-vendor β the Vulkan backend also runs on Intel graphics and older NVIDIA/AMD cards without vendor-specific drivers
- Older hardware β the oldpc build targets CUDA 11 and AVX1-only CPUs for machines that cannot run the modern default build
- CPU-only β the nocuda build runs with no GPU at all, at reduced speed relative to GPU-accelerated inference
Who Should Use KoboldCpp?
Use KoboldCpp if story writing, roleplay, or a zero-install single file matters more than raw throughput or a minimal footprint; use something else if you need multi-user production serving or the smallest possible download.
How Does KoboldCpp Compare to Ollama and Other Local Tools?
KoboldCpp's closest comparisons are other tools built on llama.cpp, plus the production-serving engines aimed at a different job entirely.
Tool | License | Best For |
|---|---|---|
| KoboldCpp | AGPL 3.0 | Zero-install, roleplay/story UI built in |
| llama.cpp | MIT | Widest hardware support, direct control |
| Ollama | MIT | Easy CLI/API, model management |
| LM Studio | Proprietary (free) | No-terminal desktop GUI |
| vLLM | Apache 2.0 | High-throughput multi-user serving |
| text-generation-webui | AGPL 3.0 | Deep UI customization, many backends |
Common Mistakes When Evaluating KoboldCpp
Most confusion comes from expecting KoboldCpp to behave like a conventional installed app, or from downloading it from an unofficial source.
Frequently Asked Questions
What is KoboldCpp?
KoboldCpp is a free, single-file local LLM application built on top of llama.cpp. It runs GGUF models with no installation step and ships with the KoboldAI Lite web interface, built for story writing and roleplay.
Who maintains KoboldCpp?
KoboldCpp is maintained by a pseudonymous developer known as LostRuins, on the repository github.com/LostRuins/koboldcpp. Releases have shipped frequently, with v1.120 current as of late August 2026.
Is KoboldCpp free?
Yes. KoboldCpp is free and open-source under the AGPL 3.0 license. There is no paid tier and no hosted product β every KoboldCpp deployment runs on hardware you control.
What license is KoboldCpp released under?
AGPL 3.0 (GNU Affero General Public License version 3), a copyleft license. This differs from the MIT license used by llama.cpp, the engine KoboldCpp is built on underneath.
Does KoboldCpp require installation?
No. It is distributed as a single executable file per platform (Windows, Linux, macOS ARM64) β download it and run it directly, with no installer, Python environment, or Docker container needed.
Does KoboldCpp support AMD GPUs?
Yes, though not through the main CUDA build. The project recommends the Vulkan option inside the nocuda build as the first choice for AMD, alongside a separately maintained rolling ROCm build for Linux.
What is KoboldAI Lite?
KoboldAI Lite is the web interface bundled inside the KoboldCpp executable. It opens automatically in a browser once the server starts, and includes editable output, memory/world-info fields, and dedicated chat, story, and adventure modes for roleplay and collaborative fiction.
Can KoboldCpp generate images or transcribe speech?
Yes. The same executable bundles Stable Diffusion image generation and Whisper speech-to-text, alongside its primary text-generation function.
Is KoboldCpp good for production or multi-user serving?
No, that is not its design goal. KoboldCpp is built for single-user local use with a rich writing interface; for multi-user, high-concurrency production serving, vLLM or NVIDIA TensorRT-LLM are the appropriate tools.
How is KoboldCpp different from Ollama?
Both run on top of llama.cpp, but Ollama focuses on a minimal CLI and model registry with no bundled UI, while KoboldCpp ships a full browser-based writing and roleplay interface (KoboldAI Lite) inside the same single executable.