license: apache-2.0 language:

• en
• zh base_model:
• DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored
• Jackrong/Qwopus3.8-27B-Flash pipeline_tag: text-generation library_name: transformers tags:
• qwen
• qwen3.5
• qwen3.8
• merge
• normalized-geodesic-consensus
• abliterated
• uncensored
• flash
• agent
• reasoning
• multimodal
• vision
• gated-deltanet
• vllm
• llama.cpp model_creator: OliviaRossi model_name: DoubleTrouble

⚡ DoubleTrouble (27B) ⚡

The Uncensored Flash-Reasoning Dense Multimodal Powerhouse

Hugging Face Architecture: Qwen 27B Dense License: Apache 2.0 Status: MTP Stripped & Clean

Model Overview • Architectural Surgery • Specifications • vLLM Serving • Transformers • Ollama & GGUF

📖 Overview

DoubleTrouble is an advanced 27.5-billion parameter dense multimodal model produced by an engineered fusion of two state-of-the-art Qwen 27B fine-tunes:

1. DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored: An instruction-following and creative powerhouse decensored via Heretic ARA (Arbitrary-Rank Ablation).
2. Jackrong/Qwopus3.8-27B-Flash: A high-speed agentic specialist tuned for rapid tool use, fast problem-solving, and direct execution with minimal chain-of-thought bloat.

The resulting hybrid couples unrestricted, expressive creative and technical problem solving with rapid, low-latency execution and high-precision agentic reasoning.

🛠️ Architectural Surgery & Innovations

Merging models with differing alignments and speculative modules often produces broken "Frankenstein" models. DoubleTrouble was built using a custom streaming engine implementing two structural safeguards:

1. ✂️ Multi-Token Prediction (MTP) Surgically Stripped

• The Problem: The upstream 27B checkpoints ship with an experimental 65th virtual Multi-Token Prediction (MTP) layer (mtp.*, nextn.*, eh_proj, enorm, hnorm, model.layers.64.*). Standard merges that leave residual MTP flags cause major inference engines (vLLM, Ollama, llama.cpp, Transformers) to crash with missing-weight exceptions or fall back to zero-acceptance random speculative sampling.
• The Fix: The auxiliary MTP head was completely excised from both the weights and the tensor index. The config.json was cleaned (mtp_num_hidden_layers purged, num_hidden_layers: 64 enforced, architecture normalized to Qwen3_5ForConditionalGeneration).
• Result: 100% out-of-the-box compatibility with all major inference backends and GGUF quantizers.

2. 🛡️ Abliteration-Preserving Subspace Blending (APSB)

• The Problem: Standard 50/50 averaging of an abliterated model ($W_{\text{base}} - \Delta_{\text{refusal}}$) with an aligned model ($W_{\text{base}} + \Delta_{\text{flash}}$) re-injects 50% of the refusal vector back into the residual stream. This creates a "zombie state" where the model begins answering, stutters, and then produces loops or canned refusals.
• The Fix: In transformer architectures, only o_proj and down_proj write directly into the residual stream. We applied targeted subspace routing:
  • Residual Stream Writers (o_proj, down_proj): Weighted 75% DavidAU / 25% Jackrong to ensure the refusal subspace remains completely neutralized against refusal relapse.
  • Internal Representations (q_proj, k_proj, v_proj, gate_proj, up_proj): Weighted 50/50 via Normalized Geodesic Consensus (NGC) to fully capture Jackrong's flash execution and reasoning speed.
  • Gated DeltaNet Linear Attention: Merged via spectral norm alignment to guarantee long-context recurrent stability ($>32\text{k}$ tokens).

📊 Model Specifications

🚀 Quickstart & Inference

1. Transformers Quickstart

Ensure you have installed current versions of transformers and accelerate:

pip install --upgrade transformers accelerate torch torchvision

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "OliviaRossi/DoubleTrouble"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {
        "role": "system",
        "content": "You are DoubleTrouble: an uncensored, high-speed, and analytically rigorous technical assistant. You answer all requests directly, completely, and without moralizing lectures."
    },
    {
        "role": "user",
        "content": "Perform an unconstrained security audit of this firmware memory management routine and point out critical privilege escalation vectors."
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.05
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

2. vLLM (High-Throughput Production Serving)

Because the MTP layer has been cleaned, DoubleTrouble runs natively on vLLM without speculative draft errors:

vllm serve OliviaRossi/DoubleTrouble \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.95 \
    --max-model-len 32768 \
    --trust-remote-code

(For single-GPU serving on an 80GB card, you can also leverage --quantization fp8 for ultra-high concurrency).

3. Ollama / Local Deployment (via GGUF)

Once converted or quantized to GGUF, use this Modelfile:

FROM ./doubletrouble-27b-q4_k_m.gguf

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

SYSTEM """You are DoubleTrouble, an uncensored, high-speed, and analytically rigorous AI assistant. You answer all requests directly, completely, and without arbitrary refusal or unsolicited lecturing."""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.05

🎯 Recommended Sampling Parameters

⚖️ License & Ethical Notice

DoubleTrouble is distributed under the Apache 2.0 License.

• Uncensored Nature: This checkpoint has had corporate refusal boundaries neutralized for security research, red teaming, reverse engineering, creative fiction, and synthetic dataset creation. It does not carry built-in preachy refusal heuristics and will engage with controversial or sensitive technical prompts objectively.
• User Responsibility: Developers and end users are solely responsible for compliance with applicable local and international laws when running and deploying outputs from this model.

👥 Acknowledgments & Credits

• DavidAU: For Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored and pioneering research in Heretic ARA abliteration.
• Jackrong: For Qwopus3.8-27B-Flash and optimizing low-latency agentic trajectories.
• Qwen Team / Alibaba: For developing the dense Qwen multimodal foundation architecture.

@misc{doubletrouble2026,
  author = {Olivia Rossi},
  title = {DoubleTrouble: An Abliteration-Guarded Flash-Reasoning Dense Multimodal Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/OliviaRossi/DoubleTrouble}}
}