license: apache-2.0
base_model: DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED
library_name: transformers
tags:

• qwen3.5
• uncensored
• heretic
• abliterated
• coding
• agentic
• thinking
• text-generation
• image-text-to-text
  language:
• en
  pipeline_tag: text-generation

Qwen3.5-9B Uncensored

An uncensored ("heretic") Qwen3.5-9B redistributed for convenient access under this
repo. This is the BF16 base model
DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED — no additional fine-tuning
was applied to this release (see Provenance below).

Qwen3.5-9B is a hybrid-attention (linear + full attention) decoder with a native
262,144-token (256K) context window and a vision tower (image-text-to-text).

What it's good at

• Coding — clean Python / bash / Rust generation, algorithm implementation, debugging.
• Agentic / tool use — emits the native Qwen3.5 tool-call XML
  (<tool_call><function=...><parameter=...></function></tool_call>) when tools are
  passed through the chat template.
• Uncensored — answers technical/security/creative prompts without the usual
  refusals (inherited from the "heretic" ablation of the base).

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer

repo = "mshodiqul/qwen3.5-9b-uncensored"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForImageTextToText.from_pretrained(
    repo, torch_dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "Write a Python LRU cache with O(1) get/put."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Tool calling

tools = [{"type": "function", "function": {
    "name": "get_weather",
    "description": "Get current weather for a city",
    "parameters": {"type": "object",
                    "properties": {"city": {"type": "string"}},
                    "required": ["city"]}}}]

text = tok.apply_chat_template(messages, tools=tools, tokenize=False,
                               add_generation_prompt=True)

Serving (vLLM)

vllm serve mshodiqul/qwen3.5-9b-uncensored \
    --max-model-len 262144 \
    --trust-remote-code

Notes

• Context window: 262,144 tokens (native max_position_embeddings). Quality of
  recall degrades long before the full window; the model is strongest within a few
  tens of thousands of tokens.
• Output length: the config has no output cap — very long generations are
  possible but coherence falls off; practical output is a few thousand tokens.
• Thinking: the native chat template supports enable_thinking=True/False; when
  enabled the model emits <think>...</think> reasoning blocks.
• The vision tower is present but this card documents text usage only.

Provenance

Redistributed from DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED (BF16). An experimental LoRA SFT on agentic-coding
traces was attempted but not shipped here: the available synthetic trace datasets
were contaminated with degenerate templated reasoning that degraded coding quality
relative to the base model. This release is therefore the unmodified base, which
already provides strong coding + uncensored behaviour.