Run KoboldCpp in Google Colab
Open the official KoboldCpp notebook from the project's GitHub repository in Google Colab, choose a GPU runtime, pick a model preset (or paste a GGUF link), and run the cell. After the model downloads it prints a public Cloudflare link that opens KoboldAI Lite or serves the API to apps like SillyTavern. The notebook sets no password, so anyone with the link can use your instance. Free sessions end when you close the tab or hit Colab's limits. We read the notebook but could not run it.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
On this page
Use the right KoboldCpp Colab notebook
Use colab.ipynb on the concedo branch of the official KoboldCpp repository; its welcome text says it is the official one. We read it on 14 November 2026 F94.
Step 1: Open the notebook and choose a GPU
Open the notebook from the repository’s badge in Colab and pick a GPU runtime. The notebook stops with “Colab did not give you a GPU due to usage limits” when it finds no GPU F94. The file’s default GPU type is a T4. We have no Google account in the lab, so we could not see which GPU the free tier gives now.
Step 2: Pick a model preset
| Gemma4 E4B Uncensored | General (with vision file) | 40960 | Untested |
| Tiefighter 13B | General | 4096 | Untested |
| Estopia 13B | General | 4096 | Untested |
| Qwen3 VL 8B | Vision | 32768 | Untested |
| Cydonia 24B | Roleplay | 10240 | Untested |
| Broken Tutu 24B | Roleplay | 10240 | Untested |
| Dans PersonalityEngine 24B | Roleplay | 10240 | Untested |
| Harbinger 24B | Instruct | 10240 | Untested |
| Muse 12B | Instruct | 49152 | Untested |
| Mistral Small 24B | Corporate | 10240 | Untested |
| GPT OSS 20B | Corporate | 32768 | Untested |
| Gemma3 12B | Corporate | 16384 | Untested |
| GLM4 9B chat | Corporate | 49152 | Untested |
Two image presets, Z-Image Turbo and Flux Klein 4B, set an image model with its text encoder and VAE (image generation). Text presets put every layer on the GPU. For your own model, tick a loader checkbox and paste a GGUF link; the checkboxes override the preset. The form also loads a vision projector, Whisper, TTS, embeddings and music models. We could not run any preset, so the last column is honest: we do not know which fit. Use the VRAM calculator with the card you were given.
Step 3: Run it and get your link
Press the two play buttons. The notebook downloads KoboldCpp from the latest release, downloads the model, and starts it with --usecuda 0, --gpulayers, your context size, --websearch, --quiet and --remotetunnel, then prints the Cloudflare address F94. Open it for KoboldAI Lite. We could not measure how long that takes.
A fallback tunnel (LocalTunnel) exists for when the Cloudflare one fails; it asks for a password on its page, which the notebook prints (read from the notebook).
Connect SillyTavern, Janitor AI and other apps
The link is a normal KoboldCpp address. For SillyTavern’s Text Completion paste it without a path; for OpenAI-compatible add /v1 (SillyTavern, other apps). Users report that very long generations can time out through the tunnel; we did not measure it.
How long it lasts and what it costs
We could not record this. The notebook says to keep the page open and check for captchas so the instance is not shut down F94. It deletes downloaded models by default (the “delete existing models” box is on), so every new session downloads again. Google sets the free limits and the paid prices and changes them: read Colab’s FAQ and its pricing page for today’s values. Saving stories to your Google Drive is an option in the form (it adds --savedatafile).
Colab’s rules
Google’s terms decide what free users may run, and they change. Read the FAQ and terms yourself. Do not try to get around its limits.
When Colab is the wrong choice
| Criterion | Colab | A rented GPU | AI Horde | Your own PC |
|---|---|---|---|---|
| Cost | Free tier with limits; paid plans | Per hour | Free | Hardware you own |
| Keeps your files | No: models are deleted each start | Depends on the disk you keep | Nothing to keep | Yes |
| Model size | Set by the GPU you are given | Set by the card you rent | Set by volunteers | Set by your hardware |
| Privacy | Runs on Google's machine | Runs on the host's machine | Volunteer machines see prompts | Stays on your PC F73 |
See cloud GPUs, the AI Horde and requirements for your own PC.
Frequently asked questions
Is KoboldCpp on Colab free?
The notebook costs nothing and Colab has a free tier with limits that Google changes. Paid plans use compute units. We did not run it, so we have no session length or cost figures.
Which models fit Colab's free GPU?
The notebook's presets name 13B models at Q4_K_S, 24B models at Q4_K_S and an 8B vision model at Q6_K, with all layers on the GPU. We could not run them, so we do not say which fit the free GPU.
Why does my Colab session disconnect?
Colab ends idle or long sessions and limits free GPUs. The notebook itself says to keep the page open and check for captchas. Google's Colab FAQ is the place for the current rules.
Can I use the Colab link with SillyTavern?
Yes, the link is a normal KoboldCpp address. Paste it without a path for Text Completion, or with /v1 for OpenAI-compatible. We did not test SillyTavern through a Colab tunnel.
Is using KoboldCpp on Colab allowed?
Read Google's Colab FAQ and terms; the rules on what free users may run change. Do not try to get around its limits.