KoboldCpp questions people ask most
Is KoboldCpp free? Yes: it is free, open-source software (AGPL-3.0) that runs AI models on your own computer, and after you have a model, chatting needs no internet connection. Below are short answers to the most common questions, each linking to the page that covers it in detail.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
Is KoboldCpp free? The short answer
Yes. KoboldCpp is free and open source, runs on your own computer, and does not need an account. The questions below are grouped by topic; use the filter to find yours. Each answer links to the page with the detail and the evidence. Our own test results carry an evidence label such as F73; everything else is upstream’s statement or a user report, and says so.
Cost and licence
Is KoboldCpp free?
Yes. KoboldCpp is free, open-source software under the AGPL-3.0 licence, with no paid tier and no account. Be careful with any site that asks you for money or shows pop-ups: it is not the project. Safety and official sources
Does it cost money to run?
Only your own electricity and hardware. A rented cloud GPU costs per hour. Cloud GPUs
Is KoboldCpp open source?
Yes, the code is on GitHub under the AGPL-3.0 licence, with some bundled parts under MIT or Apache licences. What KoboldCpp is
Is KoboldCpp the best?
It depends on what you want. It suits people who want one portable file with a story and chat page, image and voice features, and many API styles; other runners suit other goals. Alternatives and a chooser
Does Kobold AI cost money?
KoboldCpp, KoboldAI Lite and the AI Horde are free to use. The AI Horde is run by volunteers, so there can be a queue. KoboldAI, Lite and KoboldCpp
Privacy and internet
Does KoboldCpp send data or connect to the internet?
In our test on Linux a normal start with a model and one generation made no connection except to the machine itself, with no update check and no download. Some options use the network by design: a model link, HF Search, the tunnel, web search and the Horde worker. Our network test and privacy page
Is KoboldCpp offline?
Once you have the program and a model file, chatting needs no internet connection. We tested the Linux build only. What stays on your computer
Is KoboldCpp private?
Your prompts stay on your computer when you run it locally. KoboldAI Lite at lite.koboldai.net and the AI Horde are online services, and the Horde sends prompts to volunteers. Local and online Lite
Can other people reach my KoboldCpp?
By default it listens on all network interfaces, so a device on your network can reach it. Use --host 127.0.0.1 to keep it to your PC, and set a password before you share any address. Remote access and safety
Safety and content
Is KoboldCpp safe?
Download only from the project on GitHub. Windows Defender found no threats in the official Windows files we scanned. Antivirus warnings on unsigned programs do happen, and fake sites exist. Safety in detail
Why does Windows or my antivirus warn about it?
The programs are not signed with a paid certificate, so SmartScreen and some antivirus tools warn about them. Check that you downloaded from the project page and compare the file hash. Why the warnings appear
Is KoboldCpp censored?
KoboldCpp runs the model you choose. Any refusal or limit comes from that model, not from KoboldCpp. We did not test models for content. Choosing a model
Is KoboldAI down?
KoboldCpp runs on your computer, so it cannot be down. lite.koboldai.net and the AI Horde are online services that can be busy or offline. Which part is which
Computers and devices
Does KoboldCpp work on Windows 7?
We cannot confirm it: we have no Windows 7 machine. Users report a PrefetchVirtualMemory error on old systems; see the start-up page. If it will not start
Does KoboldCpp work on a Mac?
On Apple Silicon yes, with Metal as the default. There is no file for Intel Macs; they need a source build. KoboldCpp on a Mac
Can I run KoboldCpp on a phone?
On Android through Termux; there is no APK, and any APK with the name is not from the project. There is no iPhone build: use KoboldCpp on a PC from the phone. Android and Termux
Does KoboldCpp work with an AMD GPU?
Yes through Vulkan, which is inside the nocuda files. We have no AMD card, so that is upstream's statement, not our test. AMD in detail
Does KoboldCpp work on Intel Arc or an Intel GPU?
Through Vulkan, which is the only prebuilt route; there is no SYCL build. We have no Intel card to test. Backends
Can I run it without a GPU?
Yes, on the CPU, only slower. A 0.5B test model generated about 27 tokens per second on 4 CPU threads in our lab; bigger models are much slower. What your computer needs
Why does it not use my GPU?
Usually the wrong file or the wrong backend. KoboldCpp does not fail when a GPU is missing: it runs on the CPU, so check the buffer lines in the log. Fix: GPU not used
Models and features
Can KoboldCpp run safetensors or EXL2 models?
Text models only as GGUF (plus old GGML files). We gave it a safetensors file and it stopped with exit code 3. Convert the model or download a GGUF version. Getting GGUF models
Can KoboldCpp generate images?
Yes, with the built-in stable-diffusion engine. On a CPU a tiny model made a 512 by 512 picture in about 6 seconds in our test. Image generation
Can KoboldCpp generate video or music?
Upstream lists video and music models. We did not run either. Video and music
Does KoboldCpp have text-to-speech and speech input?
Yes: Whisper for speech input and TTS models such as Kokoro for spoken replies. Both ran on a CPU in our tests. Voice
Can KoboldCpp look at pictures?
Yes, with a vision model and its matching mmproj file. Without the file the model still answers, but it does not see the picture. Vision
How does KoboldCpp work?
One file bundles the engine that runs the model, a launcher window and several web pages. You start it with a model file and open its address in a browser or connect an app. How it fits together
Which model should I use?
Pick the largest model that fits your memory, then a quantization such as Q4_K_M. The calculator tells you whether it fits. Models
Why does a thinking model take so long?
It writes its reasoning first. On a 0.6B test model that cost 288 tokens against 12 with thinking off. You can switch it off. Thinking on or off
Using it
What is my API key and URL?
The URL is http://localhost:5001. There is no key unless you start KoboldCpp with --password. API tester
How do I connect SillyTavern?
Choose Text Completion with the KoboldCpp type and the address http://localhost:5001, or Chat Completion with a custom address ending in /v1. SillyTavern
Why is it slow?
Look first at where the model runs and how many layers are on the GPU; then at context size and quantization. Speed
It will not start. What now?
Read the error first, because the window can close fast. Common causes are an old CPU, a blocked file and a wrong download. Fix: it will not start
Why do replies repeat or stop short?
Repetition settings, a small output length, or a wrong chat template. Change one thing at a time. Settings and starting points
How do I keep my settings?
Save them to a .kcpps file and start with --config. Config files
How do I update KoboldCpp?
Download the new file from the project page and replace the old one; your models and config files stay where they are. Updating
What does this error mean?
Search the message in the error list. Error list
Can I run it in Google Colab?
Yes, with the official notebook. We read it but could not run it. Colab