Best Model for SillyTavern (2026): Chat, Local, Free, Embedding and Image Models
The best model for SillyTavern depends on which of its three model slots you are asking about. SillyTavern is a frontend: it writes nothing itself. It sends your chat to a chat model, it can use a second, small embedding model to search old messages, and it can use a third image model to draw scenes. People search for the best of all three, and the answers have nothing to do with each other.
For the chat model, which is the one that matters most, here is the short version. On an API, start with DeepSeek V4.1 Flash or GLM 5.3 Flash. For free, use Gemini’s free API tier or a free model on OpenRouter. Locally, pick by the memory on your graphics card. And if prose is everything and cost is not, Claude.
We can do better than opinion for the chat model, because OpenRouter publishes which models SillyTavern’s users actually run. That list is below, with prices.
Everything on this page was checked on 2 October 2026: OpenRouter’s public usage page for SillyTavern and its model list, AI Horde’s live model list, and SillyTavern’s own documentation.
A note on method. We have not run a controlled, model-by-model roleplay test. OpenRouter says SillyTavern’s users have run 450 different models, and no honest test covers that. This page uses what players actually run, what each model costs over a real session, and what SillyTavern’s documentation says. Where something is our own reading of the numbers, we say so.
New to SillyTavern? Start with the SillyTavern guide. This page assumes it is installed.
The Three Kinds of Model
| Slot | What it does | Do you need it? | Where you set it |
|---|---|---|---|
| Chat model | Writes every reply | Yes | API Connections |
| Embedding model | Turns text into numbers so old messages and files can be searched | Only if you switch on Vector Storage | Extensions → Vector Storage |
| Image model | Draws characters, scenes and backgrounds | Only if you want pictures | Extensions → Image Generation |
Most of this page is about the chat model. The other two have their own sections near the end.
What SillyTavern Users Actually Run
OpenRouter puts hundreds of models behind one API key, and it publishes usage figures for the apps that connect to it. On 2 October 2026 its page for SillyTavern said users had run 450 different models and ranked SillyTavern #4 in the Roleplay category.
These are the 20 models SillyTavern users ran the most in the last 30 days, with the price OpenRouter listed for each on the same day. Prices are in US dollars per million tokens.
| # | Model | Tokens, last 30 days | Input | Output |
|---|---|---|---|---|
| 1 | GLM 5.2 (Z.ai) | 36.3B | $0.41 | $3.99 |
| 2 | GLM 5.3 Flash (Z.ai) | 26.4B | $0.15 | $0.50 |
| 3 | Gemini 3.8 Flash (Google) | 26.3B | $0.75 | $3.75 |
| 4 | DeepSeek V3.2 | 24.5B | $0.28 | $0.42 |
| 5 | DeepSeek V4.1 Flash | 17.8B | $0.30 | $1.20 |
| 6 | DeepSeek V4 Flash 0731 | 16.3B | $0.005 | $1.28 |
| 7 | DeepSeek V4 Pro 0813 | 15.2B | $0.66 | $1.98 |
| 8 | GLM 5.3 (Z.ai) | 13.7B | $1.40 | $4.40 |
| 9 | Claude Opus 4.6 | 12.7B | $5.00 | $25.00 |
| 10 | DeepSeek V4 Flash, April version | 12.2B | $0.03 | $0.06 |
| 11 | DeepSeek V4 Pro, April version | 10.8B | $0.21 | $0.42 |
| 12 | Gemini 3.7 Flash (Google) | 10.7B | $0.75 | $3.75 |
| 13 | Gemini 3.1 Pro Preview (Google) | 9.95B | $2.00 | $12.00 |
| 14 | Claude Sonnet 4.5 | 9.94B | $3.00 | $15.00 |
| 15 | Nemotron 3 Ultra (NVIDIA) | 7.77B | free version: $0 | $0 |
| 16 | MiniMax M3 | 7.61B | $0.30 | $1.20 |
| 17 | Gemma 4 31B (Google) | 7.43B | $0.09 (also a free version) | $0.34 |
| 18 | Claude Opus 5.5 | 5.78B | $4.00 | $20.00 |
| 19 | Kimi K3 (Moonshot) | 5.28B | $2.70 | $13.50 |
| 20 | MiMo V2.5 Pro (Xiaomi) | 4.15B | $0.44 | $0.87 |
Know the limits of this list before you read it.
- It only counts OpenRouter. Local models, AI Horde and direct connections to Google, Anthropic, DeepSeek or anyone else are not in it. SillyTavern is the frontend of choice for local models, so a large part of its real use is invisible here.
- It counts tokens, not people. A few players with very large context settings weigh a lot.
- The usage page does not say whether a model was run in its free or paid version. Where a free version exists we show it.
- The match between a name on the usage page and a price on the model list is ours.
With that said, four things stand out.
1. There is no single winner. By our sum of OpenRouter’s figures, DeepSeek models are about 34% of the tokens in the top 20, GLM about 27%, Gemini about 17% and Claude about 10%. Compare that with Janitor AI, where the same kind of list is about 60% DeepSeek (we break that down here). SillyTavern users spread out more and pay more.
2. GLM is first. Two models from Z.ai hold the top two places. Outside roleplay circles GLM gets little attention, and most “best model” lists do not mention it.
3. Claude is here, and it is the old Claude. Three Claude models make the top 20. The most used is Opus 4.6, from February 2026, ahead of the current Opus 5.5. The older model is also the more expensive one: $5 and $25 per million tokens against $4 and $20. The same is true one step down, where Sonnet 4.5 ($3 and $15) is in the list and the current Sonnet 5.5 ($2 and $10) is not.
There is a likely reason, and it is our inference, not something the data says. Current Claude models reject any request that sets temperature or the other sampling values away from their defaults. Older ones accept them. Many SillyTavern presets set those values, so an old preset keeps working on an old model and fails on a new one. Some players also simply prefer how an older model writes. Either way: if you use Claude, look at which Claude your preset is calling.
4. The cheap models are very cheap. More than half of the top 20 costs under $0.50 per million input tokens.
What a Session Costs on Each Model
Here is one long session priced on the models people choose.
The turn: the model receives your character card, lorebook entries, persona and chat history — 40,000 tokens in total — and writes a 500-token reply. It is the same turn we use on our DeepSeek, Gemini and Claude model pages.
| Model | One turn | 200 turns |
|---|---|---|
| Gemma 4 31B | $0.0038 | $0.75 |
| GLM 5.3 Flash | $0.0063 | $1.25 |
| DeepSeek V3.2 | $0.0114 | $2.28 |
| DeepSeek V4.1 Flash | $0.0126 | $2.52 |
| GLM 5.2 | $0.0184 | $3.68 |
| Gemini 3.8 Flash | $0.0319 | $6.38 |
| GLM 5.3 | $0.0582 | $11.64 |
| Claude Sonnet 5.5 | $0.0850 | $17.00 |
| Gemini 3.1 Pro Preview | $0.0860 | $17.20 |
| Kimi K3 | $0.1148 | $22.95 |
| Claude Sonnet 4.5 | $0.1275 | $25.50 |
| Claude Opus 5.5 | $0.1700 | $34.00 |
| Claude Opus 4.6 | $0.2125 | $42.50 |
These figures are our arithmetic on OpenRouter’s listed prices. They are a worst case: no caching, and no thinking.
- Caching lowers the bill a lot, where you get it. Providers charge much less for input they have already seen. A long chat is mostly repeated input, so the real cost on Claude or DeepSeek’s own API can be a fraction of the number above. Caching only works if the start of your prompt stays the same from turn to turn. In SillyTavern, a lorebook entry or a vector result inserted near the top of the prompt changes it and can lose the discount.
- Thinking raises it. Many current models reason before they answer, and the reasoning is billed as output.
- A smaller context lowers it most of all. At 16,000 tokens of context, every figure is less than half.
The gap is the point. The same evening costs about $1.25 on GLM 5.3 Flash and about $42.50 on Claude Opus 4.6. That is a factor of 34. Nobody can tell you the expensive one is 34 times better. What we can say is that the choice between them is a choice about money first.
Our Picks for the Chat Model
The default: DeepSeek V4.1 Flash or GLM 5.3 Flash
Both have a 1M-token context window, both are in SillyTavern users’ top five, and both cost a dollar or two for a long session. DeepSeek V4.1 Flash is the newer of DeepSeek’s two current models; GLM 5.3 Flash is the cheap model in the family SillyTavern users run the most.
One difference matters in practice. On OpenRouter, GLM 5.3 models always think before answering, and it cannot be switched off there. DeepSeek V4.1 Flash thinks by default but lets you turn it off. For quick back-and-forth scenes, a model that can skip thinking answers sooner. For which DeepSeek model and which setting, see the best DeepSeek model for roleplay.
The mid-priced pick: Gemini 3.8 Flash
Third on the list, and the most used model from a large Western lab. It costs about $6 for the same session, and Google’s own API has a free tier for it. Two warnings from Google’s documentation matter for SillyTavern: it has deprecated temperature and the other sampling settings, and it asks you to remove prefilled model turns. Both are things SillyTavern presets commonly use. Details are in the best Gemini model for roleplay.
The prose pick: Claude
Claude is what SillyTavern’s own documentation recommends “for users who want their AI chats to have a creative, unique writing style”. It is also the most expensive family on the list by a wide margin. If you choose it, choose the current model: Sonnet 5.5 for most play, at $17 for our 200-turn session before caching, against $25.50 on the older Sonnet 4.5. Reset your sampler settings to their defaults first, or the request fails. The full comparison is in the best Claude model for roleplay.
The cheapest paid pick: Gemma 4 31B
Google’s open model is in the top 20 and costs under a dollar for the session. It has a 262K context window on OpenRouter. It is also a model you can run on your own PC with a 24GB card, which makes it a good way to find out whether a local model would be enough for you before you buy hardware.
The Free Options
“Best free model for SillyTavern” has three honest answers.
1. Gemini’s free API tier. Google’s API offers a free tier on its Flash models, with rate limits, and SillyTavern’s documentation lists Google AI Studio as having “a free tier with rate limits (Gemini Flash)”. The cost is privacy: Google says content sent on the free tier is used to improve its products.
2. Free models on OpenRouter. On 2 October OpenRouter listed 17 models with a free version. The two that appear in SillyTavern users’ top 20 are Nemotron 3 Ultra (nvidia/nemotron-3-ultra-550b-a55b:free) and Gemma 4 31B (google/gemma-4-31b-it:free). The limits are 20 requests a minute and 50 a day, or 1,000 a day once you have bought at least 10 credits in total. Every swipe is a request.
3. AI Horde. A network of volunteers who lend their graphics cards. SillyTavern connects to it “out of the box with no additional setup required”. More on it below, because it has rules of its own.
We follow the free list as a whole in the best free model for roleplay. And for how OpenRouter works inside SillyTavern, see OpenRouter for roleplay.
AI Horde: Free, With Conditions
AI Horde is the route SillyTavern users reach when they have no money and no graphics card. It is worth understanding before you rely on it.
What it had on 2 October 2026. AI Horde’s own live list showed 29 text models served by 37 workers running 72 threads between them. The most served were:
| Model | Size class | Worker threads serving it |
|---|---|---|
| Skyfall 31B v4.2 | 31B | 16 |
| Llama 3.2 1B Instruct | 1B | 10 |
| Behemoth-X 123B v2.1 | 123B | 8 |
| Impish LLAMA 4B | 4B | 5 |
| Stheno 8B v3.2 (several versions) | 8B | 7 in total |
| Magidonia 24B v4.3 | 24B | 3 |
That list changes by the hour, because it depends on who has their computer switched on. So “best AI Horde model” has a practical answer, not a fixed one: pick the largest roleplay model that has several worker threads and a short queue right now. A model with one thread and a long queue will leave you waiting, however good it is. You can select more than one model in SillyTavern, which shortens the wait.
What it costs you. SillyTavern’s documentation is direct about three things.
- Waiting. By default you connect to “the Horde’s low priority ‘guest account’”, which “means you may have to wait a long time for a reply”. Registering for a free Horde key and adding it to SillyTavern raises your priority. Sharing your own graphics card earns “kudos”, which raises it further.
- Privacy. A volunteer’s machine processes your chat. The documentation warns that a malicious worker could “log your activity (input prompts, AI responses)” and says: “When using Horde never send any personal information such as names, email addresses, etc.”
- No control over the worker. You are “at the mercy of the Worker in terms of generation wait times, AI settings, and available models”.
There is a “Trusted Workers Only” checkbox. The documentation says it limits you to workers who have been hosting for a while, and adds that they “could still be seeing prompts”.
AI Horde is still very useful for one thing: trying a local model before you download it. The models volunteers serve are the same community fine-tunes you would run at home.
The Best Local Model, by Graphics Card
A local model runs on your own PC through a backend such as KoboldCpp or Ollama. SillyTavern’s documentation sums up the trade: local models “are free to use and have no content filter”, downloads are “5-50GB each”, and “most models are not as powerful as cloud LLM APIs”.
What you can run is decided by the memory on your graphics card (VRAM). The size classes below are the ones from our local models guide. The example models are the ones Horde volunteers were serving on 2 October, which is the best public sign of what people running local roleplay models actually choose. We have not tested them against each other.
| VRAM | Size class to look for | Examples being served on AI Horde today |
|---|---|---|
| About 8GB | 7B–8B | Stheno 8B v3.2 |
| 12GB | 12B–13B | Rocinante-X 12B, mini-magnum 12B |
| 24GB | 24B–31B | Magidonia 24B, Skyfall 31B, Gemma 4 31B |
| More than one large card | 70B and above | Behemoth-X 123B |
Three rules matter more than the exact model name.
Size beats cleverness, up to what fits. A larger model at a lower quality setting usually writes better than a small one at full quality. But a model that does not fit in VRAM spills into normal memory and becomes very slow.
Roleplay fine-tunes exist for a reason. A base model is trained to be a helpful assistant. Community fine-tunes are trained to stay in character and keep a scene going. Most names on the Horde list are fine-tunes.
Local model advice goes out of date fast. Any page that names “the best local model” is describing one month. The size class is what stays true.
For setup, see SillyTavern with KoboldCpp or SillyTavern with Ollama.
The setting that makes a good local model write badly
SillyTavern builds prompts in two ways, and its documentation explains them. Chat Completion sends a list of messages. Text Completion squashes everything into “one long string” that the model simply continues.
Most local models use Text Completion. To understand where one message ends and the next begins, each needs its own instruct template, which the documentation says is “usually mentioned in the model’s documentation or download page”. With the wrong template, a good model rambles, repeats or writes your lines for you, and it looks like a bad model.
So before you decide a local model is poor, check its template. Our SillyTavern presets guide explains where those settings live.
The Best Embedding Model
This is a different kind of model with a different job. An embedding model does not write. It turns text into a list of numbers, so that SillyTavern’s Vector Storage extension can find old messages or file chunks that are similar to what you just said, and put them back into the prompt.
Start with the default. The documentation says the local source “will automatically download a compatible model”, and names it: jina-embeddings-v2-base-en. Local sources are “free and unlimited” and run on your own machine.
There are three reasons to change it.
- You roleplay in a language other than English. The default model’s name ends in “en”: it is an English model. SillyTavern even has a setting to “translate files into English before processing”, which it says is “useful when using embedding models that only support English text”. A multilingual embedding model is the cleaner fix. The documentation’s examples of other local models are
mxbai-embed-largethrough Ollama andnomic-embed-text-v1.5through llama.cpp. - You already run Ollama or llama.cpp. Then using it for embeddings too keeps everything in one place.
- You want it off your machine. SillyTavern supports API sources including OpenAI, Cohere, Google, Mistral and OpenRouter. They need a key and usually cost money, though the documentation notes “calculating embeddings is pretty cheap”.
One warning from the documentation applies whichever you pick: “Embeddings are only usable when they are retrieved using the same model that generated them.” If you change the embedding model, you must recalculate your vectors, or the search returns nonsense.
And a caution of our own. A better embedding model does not fix a memory setup that is working against you. Vector Storage moves old messages out of their place in the story, and that can make chronology worse, not better. We explain what it really does in SillyTavern memory. Read that before spending time on embedding models.
The Best Image Generation Model
SillyTavern does not draw anything itself. Its image extension sends a prompt to a source, and the source runs the image model. So the first choice is the source, and the documentation lists them. Grouped by what they cost you:
| Kind | Sources SillyTavern supports | What you need |
|---|---|---|
| Free, no setup | Stable Horde | Nothing. Volunteer-run, so expect queues |
| Free, local | ComfyUI, Stable Diffusion WebUI (AUTOMATIC1111), SD.Next, stable-diffusion.cpp, Draw Things (Mac) | A capable graphics card and a model download |
| Paid cloud | OpenAI, Google AI Studio, Black Forest Labs, Stability AI, FAL.AI, x.AI, Z.AI and others | An API key and credits |
| Subscription | NovelAI Diffusion | An active NovelAI subscription |
The practical advice:
- If you run your chat model locally on one graphics card, an image model has to share that card’s memory. Running both at once on 8GB or 12GB usually means one of them has to unload. Stable Horde or a paid cloud source avoids the conflict.
Choosing the model is half of it. For connecting a source, the generation modes and the default settings worth changing, see SillyTavern image generation.
- If you only want an occasional character portrait, Stable Horde costs nothing and is already built in.
- If you already pay for NovelAI, its image model is included in what you pay; see NovelAI pricing.
We do not rank individual image models here. That is a separate subject with its own fast-moving list, and what matters in SillyTavern is which source you can actually run. For how the dedicated AI RPG platforms handle images, see AI RPG images.
Our Recommendation
- On an API, start with DeepSeek V4.1 Flash or GLM 5.3 Flash. One to three dollars for a long session.
- For free, use Gemini’s free API tier, or a free OpenRouter model within its daily limit. Use AI Horde to try models, not for anything private.
- Locally, choose by VRAM: 8B for 8GB, 12B for 12GB, 24B to 31B for 24GB. Check the instruct template before judging a model.
- For prose above all, Claude Sonnet 5.5, with sampler settings at their defaults. Check you are not paying more for an older Claude.
- Leave the embedding model on its default unless you play in another language.
- For images, pick the source your hardware and budget allow. Stable Horde is the free start.
If a model connects but behaves strangely, SillyTavern not working lists the errors and what each one means. For how the model families compare outside SillyTavern, see the best LLM for roleplay.
Frequently Asked Questions
What is the best model for SillyTavern?
For most people who use an API, DeepSeek V4.1 Flash or GLM 5.3 Flash: both have a 1M-token context window and cost between about $1 and $2.50 for a 200-turn session before caching. Gemini 3.8 Flash is the most used model from a large Western lab among SillyTavern users on OpenRouter and has a free API tier. Claude is the choice if you care most about prose and accept a bill several times higher.
Which models do SillyTavern users actually use?
On OpenRouter, on 2 October 2026, the five most used models in SillyTavern over the last 30 days were GLM 5.2, GLM 5.3 Flash, Gemini 3.8 Flash, DeepSeek V3.2 and DeepSeek V4.1 Flash. DeepSeek models made up about 34% of the tokens in the top 20, GLM about 27%, Gemini about 17% and Claude about 10%. This counts only OpenRouter traffic, so local models and direct API connections are not included.
What is the best free model for SillyTavern?
There are three free routes. Google’s API has a free tier for Gemini Flash models, with rate limits. OpenRouter lists free models such as Nemotron 3 Ultra and Gemma 4 31B, limited to 50 requests a day, or 1,000 a day once you have bought at least 10 credits in total. AI Horde is free with no account, but it runs on volunteers’ computers, replies can be slow, and its documentation tells you never to send personal information.
What is the best local model for SillyTavern?
It depends on your graphics card’s memory. With about 8GB, an 8B roleplay fine-tune. With 12GB, a 12B model. With 24GB, a model in the 24B to 31B range, such as Gemma 4 31B or a roleplay fine-tune of that size. These are also the sizes volunteers serve most on AI Horde, where you can try a model for free before you download it.
What is the best embedding model for SillyTavern?
Start with the default. SillyTavern’s Vector Storage uses a local model called jina-embeddings-v2-base-en, which it downloads by itself and which costs nothing. It is an English model, so if you roleplay in another language, switch to a multilingual embedding model through Ollama or an API source. If you change the embedding model, you have to recalculate all your vectors.
What is the best image generation model for SillyTavern?
SillyTavern does not generate images itself. Its image extension connects to a source that you choose. The free options are Stable Horde, which needs no setup, and local tools such as ComfyUI or Stable Diffusion WebUI, which need a capable graphics card. Paid cloud sources include OpenAI, Google, Black Forest Labs and NovelAI. The best one is the one that matches your hardware and budget.
Is Claude worth it in SillyTavern?
If prose quality is what you care about most and the cost is acceptable, yes. Three Claude models are in the top 20 for SillyTavern on OpenRouter. But note which ones: the older Claude Opus 4.6 is used more than the current Opus 5.5, and it costs more, $5 and $25 per million tokens against $4 and $20. If you use Claude, check that you are not paying extra for an older model out of habit.
Why does a model work in Chat Completion but write badly in Text Completion?
SillyTavern builds the prompt in two different ways. Chat Completion sends a list of messages and the provider formats them. Text Completion sends one long text, and the model needs the correct instruct template to understand where each message starts and ends. A local model with the wrong template often rambles, repeats itself or writes your lines for you. Pick the template named on the model’s download page.
The usage figures are from OpenRouter’s public app page for SillyTavern (“Top models used by SillyTavern this month”, last 30 days), read on 2 October 2026. Prices, context windows, free models and thinking defaults are from OpenRouter’s model list on the same day, and the free-model limits from OpenRouter’s documentation. The AI Horde figures are from AI Horde’s own public status API on 2 October 2026 and change continuously. Everything quoted about API types, AI Horde, local backends, Vector Storage, embedding sources and image sources is from SillyTavern’s documentation. The family shares, the match between usage names and price-list entries, and every session cost are our own arithmetic. We have not run a controlled comparison of how these models write.
Product names and trademarks are the property of their respective owners. Arcanum is not affiliated with SillyTavern, OpenRouter, AI Horde or any model maker named here.