OliviaRossi/DoubleTrouble

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

OliviaRossi/DoubleTrouble is a 27.5 billion parameter dense multimodal model based on the Qwen 27B architecture, developed by OliviaRossi. It is engineered by fusing two Qwen 27B fine-tunes to combine uncensored instruction capabilities with low-latency flash agentic reasoning. This model excels at unrestricted, expressive creative and technical problem-solving, rapid execution, and high-precision agentic reasoning, supporting a native context length of 32,768 tokens.

Loading preview...

DoubleTrouble: Uncensored Flash-Reasoning Multimodal Model

DoubleTrouble is a 27.5 billion parameter dense multimodal model developed by OliviaRossi. It is a unique fusion of two specialized Qwen 27B models:

  • DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored: Provides uncensored instruction-following and creative capabilities through Heretic ARA (Arbitrary-Rank Ablation).
  • Jackrong/Qwopus3.8-27B-Flash: Contributes high-speed, low-latency agentic reasoning and rapid problem-solving.

This "architectural surgery" involved critical innovations to ensure stability and performance:

Key Innovations

  • MTP Stripping: The experimental Multi-Token Prediction (MTP) layer was surgically removed from the base Qwen checkpoints, ensuring 100% compatibility with major inference engines like vLLM, Ollama, and Transformers without crashes or speculative sampling errors.
  • Abliteration-Preserving Subspace Blending (APSB): A targeted merging technique was used to prevent the re-introduction of refusal vectors. Residual stream writers (o_proj, down_proj) were weighted 75% towards the uncensored base, while internal representations (q_proj, k_proj, v_proj, gate_proj, up_proj) were blended 50/50 using Normalized Geodesic Consensus (NGC) to capture flash execution speed.

Capabilities & Specifications

  • Architecture: Qwen 27B Dense (Qwen3_5ForConditionalGeneration) with 64 clean layers.
  • Parameters: 27.5 billion.
  • Context Length: Native 32,768 tokens, extendable via YaRN.
  • Multimodal: Integrated Vision Encoder (model.visual.*).
  • Performance: Designed for unrestricted, expressive creative and technical problem-solving, rapid execution, and high-precision agentic reasoning.
  • Inference: Optimized for vLLM, Transformers, and local deployment via GGUF (Ollama).

Ideal Use Cases

  • Security Research & Red Teaming: Its uncensored nature allows for direct engagement with sensitive technical prompts.
  • Creative Writing & Roleplay: Provides expressive and unrestricted creative generation.
  • Agentic Workflows: Excels in rapid tool use, fast problem-solving, and direct execution with minimal chain-of-thought bloat.
  • Technical Auditing & Code Generation: Offers high precision for technical tasks.

DoubleTrouble is distributed under the Apache 2.0 License, with users responsible for compliance with applicable laws regarding its uncensored outputs.

Finetunes
2 models