vishrutJ/SuperQwen3.8-27b-abliterated
SuperQwen3.8-27b-abliterated is a 27 billion parameter, full-BF16 multimodal language model built from Qwen/Qwen3.8-27B. It features a significant reduction in refusal behavior, corrected 'overthinking' tendencies, and verified long-context capabilities up to 262,043 tokens. This model is designed for applications requiring robust multimodal understanding and reduced content refusal without compromising core capabilities like tool use and vision.
Loading preview...
Overview
SuperQwen3.8-27b-abliterated is a 27 billion parameter, full-BF16 multimodal model derived from Qwen/Qwen3.8-27B. It has been modified using a rank-4 OBLITERATUS refusal-subspace edit to significantly reduce refusal behavior while preserving its original vision tower and MTP weights. This model does not require LoRA or inference-time adapters.
Key Capabilities
- Refusal Reduction: Achieves 0/32 refusal rate, down from 30/32 (93.75%) in the parent model, with no empty outputs.
- Corrected Overthinking: Defaults reasoning to 'medium' and includes a stop condition for 'xhigh' reasoning, passing 36/36 deterministic effort/task combinations.
- Multimodal Support: Retains full
image-text-to-textcapabilities, including vision and tool use. - Long Context: Verified to retrieve needles in prompts up to 262,043 tokens, utilizing its native context window.
- Precision: Full BF16 checkpoint with 100 declared tensors changed, ensuring integrity of vision and MTP tensors.
Good For
- Applications requiring a powerful multimodal LLM with significantly reduced content refusal.
- Tasks benefiting from robust reasoning that avoids excessive deliberation.
- Use cases demanding long-context understanding and retrieval.
- Developers seeking a full-BF16 model with original-weight quality and specific behavioral modifications.