shawnw3i/Qwopus3.8-27B-Flash-V2-abliterated

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

shawnw3i/Qwopus3.8-27B-Flash-V2-abliterated is a 27 billion parameter language model derived from Jackrong's Qwopus3.8-27B-Flash-V2. This model has been specifically modified using iterative abliteration and Arbitrary-Rank Ablation (ARA) to significantly reduce its refusal rate on harmful prompts, from 98/100 to 13/100, with minimal impact on benign behavior (0.0684 nats/token KL divergence). It is designed for local deployments, roleplay, and agent workflows where the base model's strong safety alignment causes friction, offering increased willingness to engage with a broader range of prompts.

Loading preview...

Overview

shawnw3i/Qwopus3.8-27B-Flash-V2-abliterated is a 27 billion parameter model based on Jackrong's Qwopus3.8-27B-Flash-V2. Its primary distinction is a substantial reduction in refusal behavior, achieved through a two-stage weight-level modification process: iterative abliteration followed by an Arbitrary-Rank Ablation (ARA) pass. This process directly edits the model's weights without any fine-tuning or additional training data, making the changes geometric and inspectable.

Key Capabilities

  • Reduced Refusals: Achieves an 87% relative reduction in refusals, dropping from 98/100 to 13/100 on a standard harmful-behavior benchmark.
  • Preserved Benign Behavior: Maintains nearly identical performance on benign prompts, with a first-token KL divergence of only 0.0684 nats/token compared to the base model.
  • Intact MTP Head: The multi-token prediction (MTP) speculative decoding head remains fully functional and bit-identical to the base model.
  • Inspectable Edits: All weight changes are accounted for by named layers and parameters, ensuring transparency in the modification process.

Good for

  • Local Deployments: Ideal for scenarios where the base model's strong safety alignment is too restrictive, such as roleplay or experimental use.
  • Agent Workflows: Suitable for agents that frequently encounter refusal reflexes from the base model on borderline prompts.
  • Research and Experimentation: Provides a less restrictive model for exploring language generation without excessive safety guardrails.

Caveats

This model trades off safety for willingness. Users should implement their own guardrails as needed, as it is not intended to be a safety-aligned model.

Finetunes
2 models