KoboldCpp can run with no GGUF on disk. Start it with --nomodel, open the bundled Kobold Lite, and give the OpenAI Compatible provider this API's base URL and a key. Your stories, World Info and sampler settings stay where they are; the model behind them moves to a server that can hold a 24B or a 31B. Below: the launch, each provider field, which builds belong on your card and which on the API, starting presets, and the price of an evening in Adventure mode. KoboldCpp's README and wiki and the Lite source were read on 28 October 2026.
If builds with the refusals trained out are new to you, read AI with no restrictions first; this page connects one.
What you need
- KoboldCpp 1.123.1 or newer. The releases page has
koboldcpp.exe,koboldcpp-linux-x64andkoboldcpp-mac-arm64. Skip the model download; this route loads none. - An API key made in the console. It begins with
sk-unr-and the console shows it a single time. Credit is prepaid and top-ups start at $10; a key with an empty balance connects, but its first generation comes back 402. - The base URL,
https://api.unrestrictable.ai/v1. Note the/v1at the end: one checkbox below depends on it. - A model id, taken from the model list, for instance
cydonia-24b-v4.3. Lite's Fetch button pulls that same list, so you can choose there instead.
Nothing to install if you only want to try these builds: the browser chat answers three turns a day free with no account, and the Telegram bot gives three free replies a day.
Launch KoboldCpp with no model
Either path lands in the same place: Lite in your browser on port 5001, nothing loaded, and a No Models Loaded prompt offering to connect an external API service.
From the launcher
- Start KoboldCpp and wait for the launcher window.
- Open the Loaded Files tab. Tick Allow Launch Without Models and leave Text Model blank.
- Press Launch. Your browser opens
http://localhost:5001.
From a terminal
The wiki defines --nomodel as starting the KoboldAI Lite UI by itself, with no model loaded. Add --launch and the browser opens as soon as the server is listening.
# Windowskoboldcpp.exe --nomodel --launch # Linux./koboldcpp-linux-x64 --nomodel --launch # macOS (Apple Silicon)./koboldcpp-mac-arm64 --nomodel --launchPort 5001 is KoboldCpp's default, per the wiki's --port entry. When another KoboldCpp or some other program holds it, pass a different one:
koboldcpp.exe --nomodel --launch --port 5002Save the no-model launch so it becomes one click. In the wiki's words a .kcpps file stores launcher preferences and settings; --exportconfig writes one from the arguments you passed, and --config remote.kcpps reads it back. One file for the remote session and one per local model, and switching is a double-click.
Leave --gpulayers, --contextsize, --usecuda and --usevulkan out. Each one describes how a local GGUF gets loaded. Nothing is loaded here, and the remote model's window sets the context limit.
Provider settings for the remote API
Click Yes on the No Models Loaded prompt. Lite opens Select your AI provider, the popup the hosted Lite uses too; later, AI in the top bar brings it back. Field names below come from Lite's source as read on 28 October 2026.
- Provider
- OpenAI Compatible APIAn entry in the Select your AI provider dropdown
- OpenAI API URL
- https://api.unrestrictable.ai/v1Leave Add Ver. Num ticked
- API Key (Required)
- sk-unr-live-…Made in the console; shown a single time
- Model
- cydonia-24b-v4.3Picked after Fetch, or entered under Custom
- Mode
- Chat-CompletionsNeither Text-Completions nor Responses
- Use CORS Proxy
- unticked
- Non-Standard Fields
- unticked
- Pick OpenAI Compatible API from the provider dropdown.
- Put
https://api.unrestrictable.ai/v1in OpenAI API URL. Add Ver. Num stays ticked: Lite's hint says to untick it only for an API whose URL does not end in /v1, and this one does. - Enter the key under API Key (Required).
- Press Fetch and choose an id your key can see, or press Custom and type the id as the model list spells it.
- Switch Mode to Chat-Completions. The models here answer on
/chat/completions. Text-Completions sends Lite to the older endpoint instead. - Use CORS Proxy stays off. Lite's panel notes that the key goes to the API directly unless the proxy is on. With Lite served from your own localhost, your browser talks to the API itself and no third machine sees the key.
- Non-Standard Fields stays off too. Ticked, Lite would attach Min-P, Top-K, Top-A and TFS to every request, and its own hint warns that many endpoints reject those.
- Press Ok. Lite connects, and the next turn is answered by the API.
The panel carries one warning worth reading twice: reconnecting clears custom OpenAI URLs. Switch to a local GGUF and back, and the URL and key need entering again, unless they sit in the popup's Favorites row, which stores them in your browser. On a shared computer, skip Favorites.
How the remote option fits into KoboldCpp
KoboldCpp is one executable. It runs GGUF files on your CPU or GPU and serves Kobold Lite at http://localhost:5001. The README (read 28 October 2026, release 1.123.1) also covers the reverse case: the web UI connected to an external AI provider in place of a local model, with OpenAI Compatible among the providers it lists. The documented route is the one above, --nomodel or Allow Launch Without Models on the Loaded Files tab.
That is the entire mechanism. A program already on your disk becomes the front end for models your hardware cannot load, and nothing about your stories, World Info or settings moves. This API takes the OpenAI wire format on a single key, which is why OpenAI Compatible is the provider to choose. Local models stay one launch away: quit, start again with a GGUF chosen, and Lite is back on your GPU.
The page assumes KoboldCpp on your machine. Lite users at lite.koboldai.net with no KoboldCpp installed want the KoboldAI Lite custom API guide, which walks that page's fields one by one. Scripts written against the Kobold Classic shape are covered in the KoboldAI Classic API explainer, which sets that format beside what this API returns. For choosing a build before any of this, the uncensored LLM leaderboard sorts the catalog by task and price.
Which builds stay local and which run remote
Memory decides. The wiki's table for a Q4_0 GGUF at 2048 context puts a 13B at 16 GB of system RAM or more, a 30B at 32 GB, a 65B at 64 GB, and that is before the context grows past 2K. Pushing layers to a GPU swaps RAM for VRAM; the model gets no smaller. Each of the four builds below carries the catalog's roleplay tag. The middle column is the wiki's row for that size class; prices come from the live catalog each time the page renders.
| Wire formats | |||||
|---|---|---|---|---|---|
| MythoMax L2 13B | Roleplay | $0.404 | $0.404 | 4K | |
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 | 131K | |
| Gemma 4 31B Novelist | Roleplay | $0.35 | $1.575 | 262K | |
| Uncensored AI Roleplay | Roleplay | $0.60 | $2.34 | 128K |
| Model id | System RAM, local Q4_0 GGUF (wiki) | Window in the catalog | Context Size to set |
|---|---|---|---|
mythomax-l2-13b | 13B class: at least 16 GB | 4,096 tokens | 3,584 |
cydonia-24b-v4.3 | between the 13B and 30B rows: 16 to 32 GB | 131,072 tokens | 15,872 |
gemma-4-31b-novelist | 30B class: at least 32 GB | 262,144 tokens | 15,872 |
uncensored-ai-roleplay | not a download; runs only on the API | 128,000 tokens | 15,872 |
mythomax-l2-13b: the model most Kobold users already know from a local run. Same window remote as local; what the API saves is the download and 16 GB of RAM.cydonia-24b-v4.3: the 24B roleplay tune the Lite and SillyTavern guides on this site begin with. At this size an 8 GB card can no longer take enough layers.gemma-4-31b-novelist: a story tune with a long window. Locally, 31B at the 32 GB row plus a long context asks for a workstation. Remotely, Context Size is the one limit in your hands. The uncensored Gemma page lists the rest of the family.uncensored-ai-roleplay: this site's roleplay route. No GGUF exists for it; it lives behind the API alone, so remote is the only way to run it.
A rule of thumb that holds: tunes from 7B to 12B stay local when they fit your card and you like them; 24B and above go on the API. The model list prints the window and both prices for each id, so the same check applies to whatever you were about to download.
Presets: Usage Mode, Context Size, Max Output and the samplers
On the remote provider Lite's Settings panel shows every control it shows for a local model. The difference is the provider panel's own note: Temperature, Top-P and Repetition Penalty are the only samplers sent. Min-P, Top-K, Top-A, TFS and the sampler order stay on screen and go nowhere while Non-Standard Fields is off.
| Setting | Start at | Reason |
|---|---|---|
| Usage Mode | Adventure or Chat | Four choices: Story, Adventure, Chat, Instruct. Change it per session; the connection stays as it is. |
| Context Size | table above | In the wiki's definition, the ceiling on tokens sent to the AI each turn. Each of those tokens is billed as input, so this slider sets the bill. |
| Max Output | 512 | The ceiling on reply length. A paragraph or two for adventure; go higher in Story mode. |
| Temperature | 0.9 | The wiki's local default is 0.7. Abliterated and roleplay tunes come out flat at that value; 0.9 keeps the variety. Move back toward 0.7 when a build wanders off the plot. |
| Top-P | 0.95 | The wiki says 0.92. Both work; below 0.9 the prose goes flat. |
| Repeat Penalty | 1.05 | 1.1, the wiki's value, is for a small local model stuck in a loop. On a remote 24B or bigger, 1.05 stops pet phrases and leaves names intact. |
These match the starting points the SillyTavern guide gives an abliterated roleplay build, translated to Lite's labels. Memory and World Info from the Context panel travel with every turn and count toward Context Size: a Memory of 500 tokens is 500 input tokens on each request. Lite's Autosave session (Persist Autosave Session in the wiki) brings the story and these values back on the next launch; the wiki describes it as saving story and settings for restoration the next time KoboldCpp starts.
The price of an adventure session
Worked example: 150 Adventure turns, replies of about 200 tokens. In the short column each turn sends roughly 2,000 tokens of story; in the long column 8,000, with more memory and world info behind it.
| Model | 150 turns at 2,000 context | 150 turns at 8,000 context |
|---|---|---|
| MythoMax L2 13B | $0.13 | over the window |
| Cydonia 24B v4.3 | $0.16 | $0.60 |
| Gemma 4 31B Novelist | $0.15 | $0.47 |
| Uncensored AI Roleplay | $0.25 | $0.79 |
Set that against the same session run locally: no per-token bill, but a GGUF to download, the RAM row above, and a card at full power all evening. Per-token rates for every model are on the pricing page.
Test the key before opening KoboldCpp
One request lists the ids your key can reach. It is the call Lite's Fetch button makes, it is free, and it tells a key problem apart from a KoboldCpp problem before the launcher is open.
curl https://api.unrestrictable.ai/v1/models \ -H "Authorization: Bearer sk-unr-live-…"A 200 with ids: the key works. A 401: the key is the problem, and the body says in which way:
HTTP 401{"error":{"message":"No API key was sent.","type":"authentication_error","param":null,"code":"invalid_api_key"}}HTTP 401{"error":{"message":"The API key is not valid.","type":"authentication_error","param":null,"code":"invalid_api_key"}}Troubleshooting
The launcher will not start with the model field empty
Allow Launch Without Models is unticked. Tick it on Loaded Files with Text Model blank, then Launch. On the command line --nomodel is the equivalent.
No browser window
Pass --launch, or go to http://localhost:5001 by hand. A port-in-use message means an earlier KoboldCpp is still up: quit it, or give this one a different --port.
URL and key vanished after a restart
Expected: Lite resets custom OpenAI URLs on reconnect. Store the pair in Favorites, or keep the URL in a note and paste a new key from the console.
401: the key
"No API key was sent." is an empty key field. "The API key is not valid." is a wrong, truncated or revoked key. Make a new one and paste all of it.
404 model_not_found: the model id
A misspelt Custom entry, or a build that has since left the catalog. Fetch again and pick from the list.
Fetch works, every turn 404s
Mode is Text-Completions, so Lite is hitting the text-completion endpoint. Choose Chat-Completions and reconnect.
402: the balance
The prepaid balance has hit zero. Top up in the console. A request answered 402 costs nothing.
Replies cut off mid-sentence
Max Output is too small for the mode, or Context Size plus Max Output exceeds the window. Bring Context Size down to the table value first.
Sampler changes do nothing
Only Temperature, Top-P and Repetition Penalty reach this provider. Min-P and the others belong to local models; keeping Non-Standard Fields off is correct.
The daily real-model check sends one fixed prompt to each model every day, so the build behind an id matches its label. The docs page has base URL, key and model id for other clients.
Questions
Yes. The README documents connecting the web UI to an external AI provider rather than a local model, and OpenAI Compatible is one of the listed providers. Launch with --nomodel, or tick Allow Launch Without Models on the launcher's Loaded Files tab, then choose the OpenAI Compatible API provider inside Lite.
No. KoboldCpp with no model loaded only serves the Lite page on localhost, and the model itself runs on the API. The GPU comes back into play once you launch with a GGUF chosen.
All of them: every id in the catalog speaks the OpenAI wire format. For roleplay and story work, look for the roleplay tag; the four on this page are a 13B, a 24B, a 31B story tune and this site’s own roleplay route, with live prices and windows in the tables above. See the model list.
Yes. Launch with a model chosen and Lite runs on your GPU; launch with none and set the provider for the API. One .kcpps file per setup turns each into a single click. Because Lite clears custom OpenAI URLs on reconnect, keep the URL and key in Favorites.
No. Per the wiki, --password makes clients of your own KoboldCpp server present an API key for generations. The remote API carries its own key instead, the sk-unr- one from the console, entered in Lite's API Key (Required) field.
No. These builds are abliterated or uncensored fine-tunes: the refusals were removed from the weights, so there is no jailbreak text to carry in Memory or World Info. See AI with no restrictions, explained.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.