Qwen3.5-27B-heretic LLM Model
A decensored build of Qwen3.5-27B, made with Heretic, a tool that edits the model's weights to cut refusals. Text chat with an optional thinking mode and function calling.
Overview
Qwen3.5-27B-heretic is a community variant of Qwen3.5-27B, the 27B dense model from Alibaba's Qwen team, made by coder3101 with Heretic v1.2.0. Heretic removes refusal behavior through directional ablation ("abliteration"): it finds a refusal direction inside the model and orthogonalizes selected weight matrices against it. An optimizer tunes the edit to lower refusals while keeping the model close to the original. Nothing is retrained.
The checkpoint's author reports 14 refusals on 100 test prompts, against 94 for the original model, with a KL divergence of 0.0653 from Qwen3.5-27B. The base model mixes Gated DeltaNet and gated attention layers (three to one), covers 201 languages and dialects, and is released under Apache 2.0. Qwen reports 86.1 on MMLU-Pro, 95.0 on IFEval and 72.4 on SWE-bench Verified for the original Qwen3.5-27B. The heretic card gives no benchmarks of its own.
What you can build
- Red-teaming and safety research tools that need answers the original model refuses
- Fiction, role-play and screenwriting assistants that handle mature themes
- Moderation and policy tooling that has to discuss harmful content in detail
- Agents that call your own functions, with or without step-by-step reasoning
Inputs
Input is one prompt or a messages array holding the conversation. A system prompt sets the instructions. The Thinking switch is on by default; turn it off for direct answers. Pass a user ID and session ID to carry chat history across requests.
You can set temperature (0 to 5, default 0.7), top P (default 0.95), top K (0 disables it), repetition penalty, seed, stop sequences split by semicolons and a reply length cap. Tools are OpenAI-style function definitions. Tool choice accepts auto, none or required, or names one function, and parallel calls can be allowed or blocked.
Outputs
With thinking on, the model writes its reasoning first and then the answer. With tools, the reply can be one or more function calls instead of text.
Recommended settings
Qwen's suggested sampling for the base model is temperature 1.0, top P 0.95 and top K 20 with thinking on, and temperature 0.7, top P 0.8 and top K 20 with thinking off. Set your own system prompt for the persona and limits your product needs.
Limitations
- On Wiro a request takes in about 25,600 input tokens, counting the system prompt, chat history, tool definitions and your message. The base model's 262,144-token window is not available here.
- Input over that limit is trimmed with no error, oldest turns first, then the tail of the newest message. Tool definitions are never trimmed. When they and the system prompt leave no space for your message, a context-length error comes back.
- With max tokens at 0, the reply budget is twice the prompt length, at least 2,048 tokens (8,192 with thinking on) and at most 16,384. Reasoning counts toward it.
- The base model accepts images and video. This page takes text only.
Safety & compliance
This model answers many requests that the original declines. Filter its output before it reaches end users. You remain responsible for complying with the law and Wiro's terms.
API quick start
Run Qwen3.5-27B-heretic with a single API call.
{
"prompt": "Explain the Second Law of Thermodynamics …",
"user_id": "...",
"session_id": "...",
"enableThinking": "true"
}curl -X POST "https://api.wiro.ai/v1/Run/qwen/qwen3-5-27b-heretic" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Explain the Second Law of Thermodynamics …",
"user_id": "...",
"session_id": "...",
"enableThinking": "true"
}
JSONOther models you might like
Explore similar llm & chat models.



