Qwen / Qwen3.8-27B-Obliterated
Qwen3.8-27B-Obliterated
Refusal-reduced variant of Qwen 3.8 27B for long-context chat and coding. It can emit or hide thinking traces and supports tight decoding controls.
Overview
Qwen 3.8 27B Obliterated is a text chat model derived from Qwen’s Qwen3.8-27B family. It generates an answer from your instruction, and it can optionally include a visible reasoning trace wrapped in <think>...</think> before the final response.
“Obliterated” variants are typically produced by editing weights to reduce refusal behavior. That makes the model useful for red-teaming, safety research, and uncensored prompt testing. It can also help with long-context reasoning tasks where you want deeper step-by-step work.
What you can build
- Long-document Q&A and synthesis across large context windows
- Coding help for refactors, bug triage, and design reviews
- Drafting and rewriting with tone control and style constraints
- Red-teaming prompts to test how safety behaviors change
- Agent-like planning outputs, where you ask for explicit reasoning traces
Inputs
- Your instruction as plain text. This can be a single request or a multi-part task.
- An optional “assistant rules” message as plain text. Use it to set tone, role, and boundaries.
- A toggle to show or hide the model’s thinking trace in the response.
- Optional user and session identifiers as short strings. Use them to keep chat history separated.
- Controls for randomness and diversity. You provide numeric values to make outputs steadier or more varied.
- Controls that discourage repetition. Use them when the model loops or repeats phrases.
- Output length limits. You provide minimum and maximum token budgets for the generated text.
- Stop strings as plain text. Provide one or more stop markers separated by semicolons.
- A numeric seed for repeatable runs when you want the same output pattern.
- A toggle to run with reduced-precision weights. This can reduce memory use, with possible quality loss.
- A toggle to force deterministic decoding. This turns off sampling for consistent outputs.
Outputs
The model returns a single text response. It usually contains Markdown-style formatting when helpful.
If thinking is enabled, the output may include a <think>...</think> block followed by the final answer. If thinking is disabled, it returns only the direct answer text.
Recommended settings
- For deep reasoning tasks with thinking enabled: set randomness to about 1.0, nucleus sampling near 0.95, and keep the candidate limit around 20.
- For instruction-following with thinking disabled: set randomness around 0.7, nucleus sampling near 0.8, and keep the candidate limit around 20.
- If you see loops: increase the repetition penalty slightly, and add a clear stop string.
Limitations
- Obliterated variants reduce refusal behavior. They can produce unsafe, illegal, or harmful instructions.
- Thinking output can be very long. It can consume your token budget before the final answer appears.
- Very long prompts can still degrade quality. The model may miss details or contradict earlier text.
- Low-quality inputs raise error rates. Scanned OCR text, messy logs, and broken formatting can confuse it.
- The underlying Qwen3.8-27B family is multimodal in some distributions, but this Wiro listing is text-only.
Safety & compliance
Don’t use this model to generate instructions for wrongdoing. That includes malware, weapon building, or fraud.
Treat outputs as untrusted text. Validate claims, check code, and review anything safety-critical.
Avoid sending sensitive personal data. If you must, redact it before you prompt.
If you deploy it to end users, add your own moderation, logging, and policy enforcement. This model may not self-refuse.
API quick start
Run Qwen3.8-27B-Obliterated with a single API call.
{
"prompt": "Explain the Second Law of Thermodynamics …",
"user_id": "...",
"session_id": "...",
"enableThinking": "false"
}