Mystery scene
Keep the voice
Does the reply follow the persona’s speaking style without repeating its description back to you?
English
SillyTavern model selection
What is the best model for SillyTavern?
There is no universal winner for every character card. Start with a current usage signal, then compare the same card and scene on the routes you can actually use.
See the current shortlistPopularity shows what people use on one provider network. It is not a controlled roleplay quality benchmark or a guarantee of SillyTavern compatibility.
Dated adoption snapshot
A shortlist from OpenRouter’s roleplay collection
Observed 23 September 2026
The collection says its order uses prompt and completion tokens processed through OpenRouter over the previous seven days. It measures usage across that network, not SillyTavern-only traffic or writing quality.
| Rank | Model | Share of collection usage |
|---|---|---|
| 01 | GLM 5.3 Flash | 16.7%Highest share in this snapshot; adoption signal only. |
| 02 | DeepSeek V4.1 Flash | 8.0%Second by usage in the collection at the time checked. |
| 03 | Tencent Hy4 preview | 6.5%Third; preview status may affect availability. |
| 04 | GPT-5.6 Luna | 6.2%Fourth; provider route and access can vary. |
| 05 | DeepSeek V4 Flash 0731 | 6.0%Fifth; verify the exact model ID in your backend. |
Open the OpenRouter roleplay collection ↗Trailing 7-day OpenRouter usage · roleplay collection
This dated snapshot can change. A separate modelgrep page uses a different OpenRouter roleplay-share filter and produces a different order; neither ranking is a SillyTavern-only test.
Choose for your card
Run a small comparison before a long session
Use the same character card, opening, context limit, generation settings and output length for each model. Record a few observable signals; do not treat one attractive paragraph as proof.
Repair scene
Move the scene
Does something change or become actionable, or does the answer circle around the same details?
Conversation scene
Track what happened
On the next turn, does the model remember the choice and consequences established in the scene?
These are evaluation prompts, not measured benchmark results. Repeat across several turns and note latency, cost and provider-side limits separately.
Optional comparison workflow
Keep your model test consistent
A browser with multi-model responses can help you inspect tone side by side. Tabbit can compare available models in one view; the providers, model IDs and account access still determine what is available. Tabbit is separate from SillyTavern.
These images show Tabbit Browser, not SillyTavern or measured results from this article.
FAQ
Choosing a SillyTavern model
Which model is best for SillyTavern?+
No single model is best for every card or backend. Use the dated usage list as a shortlist, then compare the same scene on the provider routes you can access.
Does SillyTavern provide the AI model?+
No. SillyTavern is a frontend that connects to an inference backend. Model access, limits, price, data handling and moderation depend on the selected provider or local server.
Does a popular model write better roleplay?+
Not necessarily. The cited usage rankings measure activity on their provider network. They do not isolate SillyTavern users or test character consistency under controlled settings.
Sources and scope