SillyTavern: what it is and what self-hosting takes

SillyTavern comes up whenever people ask for an AI girlfriend that is "free", "local" or "open source". Those words are partly true, and the part that isn't true is where most people get stuck. This guide explains what the software is, how character cards work and what a self-hosted setup really involves. We don't link to SillyTavern or to model sites, and we don't give installation steps; the aim is to help you decide whether the do-it-yourself route is worth it.

What SillyTavern is

SillyTavern is a free, open-source front-end for chatting and roleplaying with language models. It began as a fork of an earlier project called TavernAI and is maintained by volunteers. You run it on your own machine and open it in a web browser.

The key point: SillyTavern is not a model. It is a control panel. It manages characters, chat history, prompts and settings, then sends each message to a model that you connect separately. Without a model it can't answer anything. That is the main difference from apps like Candy AI or OurDream AI, where the model, images, voice and hosting all come in one package.

What it does well:

• Deep control over prompts, context length, memory tools and generation settings.
• Any backend — many cloud APIs and local model programs are supported, and you can switch between them.
• Character cards and lore — import, edit and organise characters; "world info" or lorebooks add background facts that appear when relevant.
• Group chats with several characters, and personas for yourself.
• Extensions for things like image generation, text-to-speech or translation, each of which needs its own service or model.

Character cards (角色卡)

A character card is how SillyTavern users share characters. It is usually a PNG image with the character's definition stored inside the file's metadata, so the picture and the character travel together. Plain JSON files are used too. A typical card contains:

• Name and description (appearance, background).
• Personality and the scenario the chat starts in.
• The character's first message.
• Example dialogue that shows the writing style.
• Optionally, a linked lorebook with world details.

Cards are shared on community sites and forums. Two cautions. First, a card is written by a stranger, so read the description before importing: it may include themes you don't want. Second, a card is only half of the result — the same card behaves very differently on a small local model and on a large cloud model.

What a self-hosted setup really involves

There are two ways to give SillyTavern a brain, and each has its own costs.

Option 1: a cloud model through an API key

• Accounts and keys. You register with a model provider or an aggregator, add billing and create an API key, which you paste into SillyTavern. The key is like a password tied to your card: if it leaks, someone else can spend your money.
• Cost per use. You pay for the text sent and received. Long roleplays with large context resend a lot of text every turn, so costs grow with chat length.
• Provider rules. Every provider has a usage policy, and most big ones restrict explicit content. Those rules apply to you even though the interface is open source. We don't cover ways around them.
• Privacy. Your messages go to the provider's servers, so "self-hosted" here only means the interface is on your machine.

Option 2: a local model on your own hardware

• A second program. SillyTavern needs a separate backend that loads and runs the model — common names are KoboldCpp, Ollama, LM Studio and text-generation-webui.
• A model file. You choose and download an open-weight model, often several gigabytes, usually in a compressed ("quantized") form so it fits in memory.
• Hardware. Graphics memory (VRAM) is the main limit. As a rough guide, small models of around 7–8 billion parameters run on a GPU with about 8 GB of VRAM; mid-size models want 12–24 GB; the large models that write best need far more, or multiple GPUs. Computers with a lot of unified memory can run bigger models, more slowly. On a CPU alone, replies can take a long time.
• Quality trade-off. What fits on a home computer is usually smaller than the models behind commercial apps, so expect more repetition and forgetfulness unless you invest in hardware.
• Privacy. This is the real advantage: chats can stay entirely on your machine.

Time

Plan on an evening or more for a first working setup, then ongoing tuning: prompt formats that match your model, context size, generation settings, updates to SillyTavern and the backend, and finding cards you like. For many hobbyists that tinkering is the fun part. If you just want to chat tonight, it is not.

SillyTavern vs a ready-made AI girlfriend app

Whichever route you take, adult content is for adults only. Mainland Chinese law prohibits obscene content; follow the law where you live. Our AI girlfriend safety guide covers billing, data and deletion for hosted apps.

Don't want to self-host?

If what you want is the result — a character with memory, photos and voice — without managing models, these live partners run everything in the browser:

• Candy AI — the simplest start: $3.99/mo on a 12-month plan that renews at the same price, Girls / Anime / Guys, images, video and voice, 13 interface languages.
• OurDream AI — closest to SillyTavern's roleplay depth: millions of user-made characters, group chats, images, video up to 60 seconds and voice calls.
• GirlfriendGPT — for people who like choosing models: a huge community bot library, memory you can view, edit and pin, a model choice, and a Simplified Chinese interface.
• Kupid AI — 15 free messages, Simplified Chinese, and passes that never auto-renew, paid by card, PayPal or crypto.

More options: uncensored AI chat apps and free AI girlfriend apps.

FAQ

Is SillyTavern free?

The software is free and open source. The model behind it is a separate matter: a cloud API usually charges per use, and a local model costs nothing per message but needs capable hardware and electricity.

Does SillyTavern include an AI model?

No. SillyTavern is only the interface. It sends your messages to a model you connect — either a cloud service through an API key or a model running on your own computer through a separate program.

What is a SillyTavern character card?

A file, usually a PNG image, that carries a character's definition inside it: name, description, personality, scenario, first message and example dialogue. You import the card and the character is ready to chat.

Do I need a GPU for SillyTavern?

Not for SillyTavern itself, which is light. You need a strong GPU (or a computer with lots of fast memory) only if you want to run the language model locally. With a cloud API, an ordinary computer is enough.

Is SillyTavern available in Chinese?

The interface has community translations, including Chinese, and Chinese-speaking users often call it 酒馆. Whether replies come in good Chinese depends on the model you connect.