veldierin/Qwopus3.8-27B-Flash-V2-Apostate-Uncensored

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

veldierin/Qwopus3.8-27B-Flash-V2-Apostate-Uncensored is a 27 billion parameter uncensored language model derived from Jackrong/Qwopus3.8-27B-Flash-V2. It was created using the 'apostate diode' method, which selectively removes refusal behaviors by modifying only 0.0019% of the model's parameters. This model retains the original's general capabilities while significantly reducing refusals to harmful prompts, making it suitable for applications requiring less restrictive content generation. It supports a 32768 token context length and integrates with standard inference engines like vLLM and llama.cpp without special runtime hooks.

Loading preview...

Overview

This model, veldierin/Qwopus3.8-27B-Flash-V2-Apostate-Uncensored, is an uncensored variant of the 27 billion parameter Jackrong/Qwopus3.8-27B-Flash-V2 model. It was created using the apostate diode method, a technique that identifies and subtracts refusal directions within the model's weights. This process modifies only a tiny fraction (0.0019%) of the parameters, resulting in a standard checkpoint that loads and runs like any other model, without requiring special runtime hooks or adapters.

Key Capabilities

  • Significantly reduced refusal rate: The model refuses only 14 out of 94 harmful prompts, compared to 92 out of 94 for the base model, and 8 out of 100 on a standard 100-prompt set (vs. 98/100 for base).
  • Preserved general ability: Extensive testing across knowledge tasks (ARC, HellaSwag, Winogrande, BoolQ, OpenBookQA, PIQA) and a six-category capability test (knowledge, maths, truth, instruction-following, code, tool calls) shows no detectable degradation in general performance compared to the base model.
  • Standard inference compatibility: Functions seamlessly with AutoModelForCausalLM, vLLM, llama.cpp, Ollama, and LM Studio.
  • Fast speculative decoding: Ships with the original MTP (speculative draft) block, enabling high decode and prefill speeds (e.g., 130 tokens/second decode, 1220 tokens/second prefill with Q5_K_S quant).

Good for

  • Applications requiring a less restrictive or uncensored language model for content generation.
  • Use cases where maintaining the base model's general capabilities (reasoning, instruction-following, code, maths) is crucial, but refusal behaviors need to be minimized.
  • Developers seeking a model that integrates easily into existing LLM inference pipelines without complex setup.

Limitations

  • Uncensored nature: This model will answer harmful requests and has no added safety layers. Users are responsible for its deployment.
  • Partial removal: Some harmful prompts (14/94) may still elicit refusals.
  • Untested vision: While vision weights are byte-identical, end-to-end vision capabilities have not been tested on this specific build.

Finetunes
1 models