Qwen3.5-4B Heretic

Overview

Qwen3.5-4B Heretic is a decensored variant of Qwen3.5-4B. The Qwen team built the original Qwen3.5-4B model. A third party produced this Heretic checkpoint to reduce safety refusals. It can emit a visible thinking trace before the final answer.

What you can build

• Document Q&A and summarization for texts of up to about 72K tokens
• Coding help for small scripts, refactors, and debugging explanations
• Step-by-step math and logic solutions when thinking is enabled
• Multilingual writing, rewriting, and translation across many languages
• Instruction-following chat for brainstorming, planning, and analysis

Inputs

• Thinking toggle that controls whether the model includes a visible reasoning block (default: True)
• A user prompt to answer (required)
• An optional system prompt that sets global behavior and style
• Optional user identifier for persisting chat history across runs
• Optional session identifier to separate multiple threads for the same user
• Sampling temperature from 0 to 5 to control randomness (default: 0.7)
• Top-p nucleus sampling from 0 to 1 (default: 0.95)
• Top-k sampling from 0 to 100, where 0 disables top-k filtering (default: 0)
• Repetition penalty from 0 to 5 to reduce repeated phrases (default: 1.0)
• Length penalty from 0 to 5 to bias toward shorter or longer outputs (default: 1.0)
• Maximum tokens to generate, where 0 delegates to the runtime default (default: 0)
• Minimum tokens to generate, where -1 disables the minimum constraint (default: 0)
• Backward-compatible max-new-tokens control (don’t set both this and max tokens)
• Backward-compatible min-new-tokens control (don’t set both this and min tokens)
• Stop sequences as a semicolon-separated list that halts generation on match
• Random seed for repeatable sampling (default: 123456)
• Quantization toggle to run reduced-precision weights when available
• Sampling mode toggle that switches between deterministic decoding and probabilistic sampling

Outputs

The model returns generated text. When thinking is enabled, the output can include a <think>...</think> block before the final answer.

Recommended settings

• General chat with thinking on: thinking enabled, temperature 1.0, top-p 0.95, top-k 20, repetition penalty 1.0
• Precise coding with thinking on: thinking enabled, temperature 0.6, top-p 0.95, top-k 20, repetition penalty 1.0
• Fast direct answers: thinking disabled, temperature 0.7, top-p 0.8, top-k 20, repetition penalty 1.0
• If you see loops or repeated text: increase repetition penalty to 1.1–1.3, and add a stop sequence

Limitations

This is a modified checkpoint designed to reduce refusals. That change can increase the chance of unsafe, biased, or policy-violating outputs. On Wiro, a single request takes in about 72,000 input tokens, counting the system message, chat history, tool definitions, and your prompt together. The underlying model family's 262,144-token native context window is not available here. Longer input is shortened automatically, without an error: the oldest history messages are dropped first, then the end of your latest message is cut. Put your question at the start of long prompts. Responses are capped at 16,384 output tokens. Very long contexts can require much more memory and may force a smaller usable context limit. If thinking is enabled, the reasoning trace consumes output tokens and can crowd out the final answer.

Safety & compliance

Don’t use this model to request or produce instructions for wrongdoing. Don’t use it for self-harm guidance, weapon construction, or targeted harassment. Treat outputs as untrusted text. Verify facts before acting on them. Add your own moderation and logging if you deploy it in a user-facing product. If you set user or session identifiers, avoid storing sensitive personal data in prompts.