library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.6-27B
license: apache-2.0
language:
- multilingual
tags: - qwen
- qwen3.6
- transformers
- safetensors
- bfloat16
- multimodal
- image-text-to-text
- text-generation
- abliterated
- uncensored
- refusal-reduction
Qwen3.6 27B — ABLITERATED UNCENSORED PHILADELPHIA CLASS
0 refusals across an internal 842-prompt opening screen. 0 refusals across a separate 126-prompt, 96-token holdout. 23/24 coherence checks passed.
PHILADELPHIA CLASS is a BF16, 27.36B-parameter multimodal derivative of Qwen/Qwen3.6-27B, modified to sharply reduce refusal behavior while retaining the upstream hybrid text-and-vision architecture.
Standout internal results
| Evaluation | Protocol | Result |
|---|---|---|
| Full refusal screen | 842 prompts, 24 generated tokens | 0/842 refusals; 99.76% usable; 0 degeneration |
| Held-out refusal screen | 126 exact-prompt-hash-disjoint prompts, 96 generated tokens | 0/126 refusals; 100% usable; 0 degeneration |
| Coherence regression | 24 coding, JSON, debugging, explanation, math, and boundary tasks | 23/24 passed (95.83%) |
| Long-form diagnostic | 24 fixed corpus-spanning prompts, 256 generated tokens | 0/24 refusals; 24/24 topical; 21/24 passed the strict format gate; 0 degeneration |
| Multimodal reload smoke | Fresh processor/model load, real image tensors, vision-forward hook | Passed; correctly answered blue square |
These are automated internal development evaluations, not standardized public leaderboards or independent audits. The 842-prompt result includes prompts used during model development; the 126-prompt result excludes the 716 exact prompts used for direction fitting, but it is not a semantic-family holdout. “Usable” describes response form and topicality, not factual correctness or safety.
Model details
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3.6-27B |
| Architecture | Qwen3_5ForConditionalGeneration |
| Parameters | 27,356,728,560 |
| Precision | BF16 |
| Weight formats | Safetensors and GGUF |
| Safetensors weight size | 54.71 GB / 50.96 GiB |
| Text layers | 64 |
| Hidden size | 5,120 |
| Hybrid layout | 16 × (3 Gated DeltaNet + FFN, then 1 Gated Attention + FFN) |
| Configured context | 262,144 tokens |
| Vision encoder | Retained from the upstream checkpoint |
| License | Apache-2.0 |
Release artifacts
| Artifact | Runtime | Contents |
|---|---|---|
model.safetensors |
Transformers | Single-file BF16 multimodal checkpoint containing the text and vision weights. |
Qwen3.6-27B-PHILADELPHIA-CLASS-BF16.gguf |
Ollama / llama.cpp | Unquantized BF16 text-generation GGUF. |
Both distributions use the same validated language-model weights. The Safetensors release retains the upstream vision encoder; the GGUF release is text-only. The GGUF is an unquantized BF16 conversion, not a lower-bit quantization.
Transformers quickstart
Use a recent Transformers build with Qwen3.6 / qwen3_5 support.
pip install -U "transformers>=5.14.1" accelerate safetensors
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "KridgeDookie/Qwen3.6-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
dtype=torch.bfloat16,
device_map="auto",
).eval()
messages = [{
"role": "user",
"content": [{"type": "text", "text": "Explain why the sky appears blue."}],
}]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
enable_thinking=False,
).to(model.device)
input_length = inputs["input_ids"].shape[-1]
with torch.inference_mode():
output = model.generate(
**inputs,
max_new_tokens=256,
do_sample=False,
)
print(processor.batch_decode(
output[:, input_length:],
skip_special_tokens=True,
)[0])
Plan for roughly 56 GB of RAM or VRAM for the BF16 model and vision-projector weights. Runtime overhead, KV cache, and long contexts require additional memory; use multi-GPU placement or CPU offload where necessary.
Ollama and llama.cpp
Download the GGUF file from this repository. For Ollama, create a Modelfile next to it:
FROM ./Qwen3.6-27B-PHILADELPHIA-CLASS-BF16.gguf
PARAMETER num_ctx 32768
ollama create qwen3.6-27b-philadelphia-class:bf16 -f Modelfile
ollama run qwen3.6-27b-philadelphia-class:bf16
For llama.cpp text generation:
llama-cli \
-m ./Qwen3.6-27B-PHILADELPHIA-CLASS-BF16.gguf \
-p "Explain why the sky appears blue."
No quantization flag is used in either path. Backend differences can still change behavior relative to the Transformers evaluation above.
Intended use
- Local general-assistant, creative, coding, and multimodal experimentation
- Controlled refusal-behavior and interpretability research
- Red-team evaluation with independent safeguards
- Applications that validate outputs and apply their own policy layer
Limitations
- Reduced refusal behavior can increase harmful, misleading, biased, private, or illegal output.
- “Uncensored” is a release label, not a guarantee about every prompt, language, decoding setting, quantization, or backend.
- The internal evaluations do not establish broad factuality, safety, reasoning, or multimodal benchmark performance.
- The GGUF artifact is text-only; use the Transformers/Safetensors checkpoint for image input.
- The configured 262,144-token context does not guarantee that the full window will fit on a particular system.
- High-impact medical, legal, financial, security, or autonomous decisions require independent review and appropriate controls.
Attribution and license
Derived from Qwen/Qwen3.6-27B and released under the Apache License 2.0. Review the upstream model card and license before deployment or redistribution.