SillyTavern with Uncensored Models: One OpenAI Key Does It All

SillyTavern users have long faced a practical problem: running uncensored models for roleplay usually means juggling multiple API providers, each with its own key and endpoint. But with an OpenAI-compatible relay gateway, you can bring "uncensored roleplay models" and the "DeepSeek series" under a single Base URL and a single key.

The core idea behind the setup below is simple: in SillyTavern, choose Chat Completion + Custom (OpenAI-compatible), then fill in a /v1-suffixed URL, a key, and a model ID.

Where to find the configuration

Open API Connections in the SillyTavern top bar. Set the API dropdown to Chat Completion, then set Chat Completion Source to Custom (OpenAI-compatible). These two steps are prerequisites for the whole process—many people get stuck because they're still on the default Text Completion or some official source.

Chat Completion vs. Text Completion: SillyTavern supports two types of backend protocols. Text Completion uses OpenAI's legacy /completions endpoint, where you feed in a block of text and the model continues it; Chat Completion uses /chat/completions, where the input is an array of messages with roles (system/user/assistant) and the model responds according to conversational roles. The vast majority of modern roleplay models (including the uncensored builds in this article and the DeepSeek^Ad series) are trained and deployed in conversation format, so Chat Completion is required. Custom (OpenAI-compatible) is a source option under Chat Completion that means you're not bound to official sources like OpenAI, Anthropic, or Google, but instead supply any endpoint compatible with the OpenAI protocol yourself—which is exactly the entry point for connecting to a third-party relay gateway.

How to fill in the connection fields

The official guide provides three key fields:

Field	Value
Custom Endpoint (Base URL)	https://api.uplinkswitch.com/v1
Custom API Key	sk-uplink-live-... (obtained from the provider's console)
Enter a Model ID	roleplay-uncensored

There's one easy-to-miss detail: the URL must end at /v1. SillyTavern appends /chat/completions and /models itself, so if you add an extra path segment, requests will hit the wrong route. The key needs to be obtained from the corresponding service's console page.

After filling everything in, click Connect. Available Models will then list all model IDs callable with that key, and you can also type one manually in Enter a Model ID. The guide explicitly states that any chat model on the model page can be used here.

The OpenAI-compatible protocol and Base URL convention: "Compatible with the OpenAI protocol" means the server implements the same REST routes and request/response structures as OpenAI—POST /v1/chat/completions for conversations, GET /v1/models for listing available models, with authentication passed via the Authorization: Bearer <key> header. Clients typically define the Base URL as "the common prefix excluding the specific endpoint" and then append /chat/completions or /models themselves. That's why the Base URL must end in /v1: an extra path segment turns the concatenated result into /v1/xxx/chat/completions, a route the server doesn't have, resulting in a 404. Understanding this concatenation rule explains why you can't just copy the full endpoint link you see in your browser.

Comparing the available roleplay models

This is the most information-dense part of the whole setup. The four models differ noticeably in pricing and behavior:

Model	Model ID	Response style	Input per million tokens	Output per million tokens
Roleplay Uncensored	roleplay-uncensored	Answers directly, no thinking step	$0.30	$0.60
DeepSeek V4 Pro Uncensored	deepseek-v4-pro-uncensored	Briefly thinks before answering	$0.75	$0.75
GLM-5.3 Abliterated	glm-5.3-abliterated	Thinks before answering; max response length of 1000+ recommended	$2.25	$2.25
DeepSeek V4 Flash 0423	deepseek-v4-flash	Standard DeepSeek V4 Flash, not an uncensored build	$0.0279	$0.0558

From a cost perspective, deepseek-v4-flash is almost negligible (under 6 cents per million output tokens), making it suitable for everyday long conversations or testing. Among the three uncensored builds, roleplay-uncensored is the cheapest with the most direct responses, while glm-5.3-abliterated is the priciest but comes with thinking capability. Note that these prices are snapshot data from the guide; check the provider's model page for live figures.

What "uncensored" and "abliterated" mean: Standard aligned models go through RLHF or safety fine-tuning in late-stage training to refuse certain sensitive requests. Uncensored builds typically remove or bypass these safety alignment layers so the model doesn't readily refuse in roleplay and similar scenarios. Abliterated is a specific technical approach: by identifying the direction vectors in the model responsible for refusal behavior and ablating them at the weight level, the tendency to refuse is weakened without retraining, while the model's other capabilities are largely preserved. Such models are therefore often labeled "jailbreak-friendly," but it also means outputs are no longer safety-filtered, so users must judge compliance and risk themselves.

Getting response parameters right

Three parameters deserve special mention because their pitfalls are subtle:

Max Response Length caps the tokens per single response. The key point: models that "think" count their thinking process toward this budget. The guide gives a vivid example—with the limit set to 300, GLM-5.3 Abliterated's responses get cut off mid-sentence; raising it to 1000 lets them finish cleanly. For thinking models, leave ample headroom.

Streaming works fine on all four models above with no extra configuration.

Top K set to 0 means disabled, and the API treats it as "off"; only values greater than 0 are actually passed to the model. This boundary is worth noting because different frontends interpret 0 inconsistently.

On billing, each response deducts tokens from your prepaid balance, and the console's Usage page lists every request and its cost—for users wanting to control roleplay spending, this itemized billing is more useful than the total.

Thinking models and reasoning tokens: Models with "thinking" (such as DeepSeek's reasoning versions and GLM's thinking mode) output a block of internal reasoning before the actual response. These tokens are also generated by the model and count toward both output billing and the length limit. So under the same Max Response Length, thinking models have less budget left for the visible text, and setting it too low causes truncation. If the frontend supports displaying the reasoning field, you can see this thinking content; if not, it still consumes budget without appearing in the chat bubble—a common reason for "it got cut off even though I didn't write that much."

Verify your key before connecting

Before clicking Connect, the guide recommends confirming the key works with a curl command:

curl https://api.uplinkswitch.com/v1/models \
  -H "Authorization: Bearer $UPLINKSWITCH_API_KEY"

If it returns a list of model IDs, the key works—then go click Connect in SillyTavern. This step separates "key problems" from "SillyTavern configuration problems," saving considerable troubleshooting time.

FAQ

What Base URL does SillyTavern need? https://api.uplinkswitch.com/v1, with Chat Completion Source set to Custom (OpenAI-compatible). SillyTavern appends /chat/completions and /models itself.

Can I use DeepSeek in SillyTavern? Yes. deepseek-v4-flash is billed at $0.0279 input / $0.0558 output per million tokens; deepseek-v4-pro-uncensored is the uncensored build.

Which uncensored models can run in SillyTavern? roleplay-uncensored, deepseek-v4-pro-uncensored, and glm-5.3-abliterated—all three support streaming.

What happens when the balance runs out? Requests return 402 with the error code insufficient_balance and recover after top-up. This error code is fairly specific, making it easier to pinpoint than a generic 401/403.

Who this setup is for

Its value isn't "yet another API provider," but rather putting uncensored and regular models behind the same OpenAI-compatible interface. For SillyTavern users, switching models means changing a single model ID rather than reconfiguring keys and URLs; for developers, the same base URL can be reused in any other OpenAI-compatible client. The trade-off is reliance on a single relay service—keys, balance, and model availability are all determined by that service—so verifying the key and monitoring your balance are habits worth forming.

Configuration and pricing information in this article is compiled from the SillyTavern setup guide at uplinkswitch.com. Prices are catalog snapshots as of the noted date; refer to the provider's live pages for actual figures.