What KoboldCpp is, and what it can do on your PC
KoboldCpp is a free, open-source program (AGPL-3.0) by LostRuins (known as Concedo) that runs AI language models in the GGUF format on your own computer, using the CPU, an NVIDIA, AMD or Intel GPU, or Apple Silicon. It is a single file with no installer, opens a chat and story interface in your browser, and serves the same model to other apps through KoboldAI, OpenAI-compatible and Ollama-compatible APIs. It can also generate images, read images, transcribe speech and speak text.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
- gha-w25: Windows Server 2025 test machine · 4 vCPU · 16 GB · no GPU
- gha-w22: Windows Server 2022 test machine · 4 vCPU · 16 GB · no GPU
- gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
- gha-mac15: macOS 15 test VM (Apple Silicon) · 3 vCPU · 7 GB · paravirtual Metal
On this page
What is KoboldCpp? In one picture
Everything talks to one program on one port. By default that is localhost:5001 F05.
- You, or an appthe KoboldAI Lite page in your browser, SillyTavern, a coding tool, your own script
- KoboldCpp, port 5001one program answering the KoboldAI, OpenAI, Ollama and other APIs at once
- A backendCPU, CUDA (NVIDIA), Vulkan (AMD, Intel, NVIDIA) or Metal (Apple)
- Your model filea GGUF file you downloaded, loaded into memory
What you get in the single file
One download, no installer. The bundled web page, KoboldAI Lite, appears at the same address F31:
| Text generation | Runs GGUF language models on CPU or GPU | 1.41 (GGUF) | every test we ran |
| KoboldAI Lite | Chat, story, instruct and adventure modes in your browser | embedded; v353 in 1.124.2 | F31, F61 |
| Several APIs at once | KoboldAI, OpenAI, Ollama and A1111 routes on one port | OpenAI chat 1.46; Ollama 1.79 | F23, F72 |
| Image generation | Stable Diffusion and newer models | 1.60 | untested |
| Vision | Reads images with a matching projector file | 1.61 | F27 |
| Speech to text | Whisper transcription | 1.67 | F27 |
| Text to speech | Several voice models | 1.82 | untested |
| Embeddings | Vectors for search | 1.87 | F27 |
| Music generation | Models such as ACE-Step | 1.109 | untested |
| Web search, MCP tools, agent | Search, tool servers and a command-line agent | 1.81, 1.106, 1.122 | untested |
Each feature has its own guide. Start with getting started for text, then the API, images and voice guides.
Where it came from
It began in March 2023 as llamacpp-for-kobold by LostRuins (Concedo), a fork of llama.cpp that added a KoboldAI-style API. It was renamed KoboldCpp in version 1.0.8 in April 2023. It is built on llama.cpp and inspired by KoboldAI, whose original Python program now calls itself legacy and points users here. The project is on GitHub. It is released often:
What it needs
- A computer: 64-bit Windows, Linux or an Apple Silicon Mac. There is no build for Intel Macs F01.
- Memory for the model: the model file plus a little more, in RAM or in a graphics card’s VRAM. The calculator tells you.
- A CPU with AVX2 for the fast builds; older ones work with slower builds F03 F15.
More in requirements.
What it does not do
- It is not a cloud service. There is no account, and it runs only while your computer runs it.
- There is no official mobile app. On Android it can be built in Termux (Android).
- It does not train or fine-tune models.
- Models are not included. You download GGUF files yourself (Hugging Face). Answer quality depends on the model and your hardware.
Is it private, offline and uncensored?
Private and offline: your prompts stay on your computer unless you turn on a tunnel, the AI Horde or web search. After the program and a model are downloaded it works with the network off; our test saw no connection while generating F73. We have not yet started it with the network blocked on Windows. Censored or not is a property of the model, not of KoboldCpp: KoboldCpp does not add filters. Different models behave differently, so read a model’s card (models).
Who uses it for what
- Chat and an assistant: getting started.
- Stories and roleplay, in the bundled page or in SillyTavern: Kobold Lite, SillyTavern.
- A backend for other apps and coding tools: the API, integrations.
- Images and voice: image generation, voice.
KoboldCpp compared in one table
| Criterion | KoboldCpp | Ollama | LM Studio | llama.cpp | text-generation-webui |
|---|---|---|---|---|---|
| Licence | AGPL-3.0 | MIT | Proprietary | MIT | AGPL-3.0 |
| Single file, no install | Yes works F01 | Installer and service | Installer | Release folder | Portable builds |
| Built-in web UI | Yes (Lite) | No (CLI and apps) | Desktop app | Basic web UI | Yes |
| OpenAI-compatible API | Yes works F72 | Yes | Yes | Yes | Yes |
| Image generation | Yes, built in | No | No | No | Through extensions |
| GUI launcher | Yes | No | Yes | No | Yes |
The other columns are from each project’s own pages (checked 13 Nov 2026), not our tests. For a longer comparison see alternatives.
Where to go next
Answer all three to see your reading order.
Frequently asked questions
Is KoboldCpp free?
Yes. It is open-source software under the AGPL-3.0 licence, and downloading and using it costs nothing. The AI models you load are separate files with their own licences, and some are free and some are not.
Does KoboldCpp need the internet?
Not to generate text. On our Linux test machine, starting it with a model and generating a reply made no connection to any address except the machine itself. You need the internet once to download the program and a model, and for the features that use it by design: a model URL, web search, a remote tunnel and the AI Horde.
Is KoboldCpp safe?
The official download is the GitHub releases page of the project. Fake sites exist, and the project's own README warns that koboldcpp.com is malicious. See the safety page for how to check a download.
Can KoboldCpp run on a phone?
There is no official app. On Android it can be built and run in Termux, which takes a while and needs a capable phone. There is nothing for iPhone.
Is KoboldCpp the same as KoboldAI?
No. KoboldAI is the original Python program that inspired it, now unmaintained and pointing users to KoboldCpp. KoboldCpp is a separate program built on llama.cpp, and its web page is called KoboldAI Lite. The Kobold explained page maps the whole family.