Qwen / Qwen3.8-27B-Uncensored
Qwen3.8-27B-Uncensored
Chat model based on Qwen 3.8 27B with refusal behavior reduced through direction removal. Its 262k context fits long documents, code, and chats.
Overview
Qwen 3.8 27B Uncensored is a 27B-parameter chat model based on Qwen/Qwen3.8-27B by Qwen. It generates assistant text from your instructions and conversation context. You can switch a “thinking” mode on or off to trade deeper reasoning for shorter answers. This uncensored variant reduces refusal behavior by removing refusal directions, while keeping the base skills.
What you can build
- Long-document Q&A from pasted policies, specs, or research notes
- Code generation and debugging help for scripts, services, and tooling
- Agent-style task plans with step lists, checklists, and acceptance criteria
- Multi-turn troubleshooting chats for logs, configs, and incident notes
- Writing help for technical docs, emails, and structured reports
- Brainstorming and outline drafting for product and research work
Inputs
- Your main instruction or question as plain text.
- An optional system message that sets the assistant’s role, tone, and rules.
- A toggle to enable or disable “thinking” style outputs.
- Optional user and session identifiers as text, used to keep separate chat histories.
- Randomness controls for generation, including temperature and probability sampling limits.
- A repetition control that penalizes repeated phrases and loops.
- A length control that nudges the model toward shorter or longer completions.
- Optional minimum and maximum output token limits to control response length.
- Optional stop phrases as a semicolon-separated list. Generation stops at the first match.
- An optional integer seed for repeatable outputs when sampling is enabled.
- An optional quantization toggle to trade quality for lower memory use.
- A sampling toggle that switches between deterministic output and randomized sampling.
Outputs
The model returns a single assistant response as UTF-8 text. When thinking is enabled, the response may include a dedicated reasoning block before the final answer.
Recommended settings
Use these settings as a starting point, based on the official Qwen 3.8 sampling guidance.
- For harder reasoning and coding: keep thinking enabled, set temperature to 1.0, top-p to 0.95, top-k to 20, and repetition penalty to 1.0.
- For direct answers and chat: disable thinking, set temperature to 0.7, top-p to 0.80, top-k to 20, and repetition penalty to 1.0.
Limitations
- “Uncensored” does not mean “no refusals.” Refusal behavior is reduced, not fully removed.
- Total prompt plus output must fit the model’s native 262,144-token context window.
- Very long chats can still degrade recall and increase hallucinations.
- Quantization can change tone and accuracy, especially on edge cases.
- Low-quality input hurts results. Ambiguous prompts, pasted junk text, or inconsistent formatting can confuse the model.
- This listing is text-first. Image and video understanding need a multimodal input path.
Safety & compliance
This model can be more willing to answer sensitive requests than aligned chat models. You must still follow platform rules and all applicable laws. Don’t use it for instructions that enable wrongdoing, violence, self-harm, or illegal access. If you’re building a user-facing app, add your own moderation, logging, and abuse controls.
API quick start
Run Qwen3.8-27B-Uncensored with a single API call.
{
"prompt": "Explain the Second Law of Thermodynamics …",
"user_id": "...",
"session_id": "...",
"enableThinking": "true"
}