DreamFast/Qwen3-VL-4b-Heretic

VISIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DreamFast/Qwen3-VL-4b-Heretic is a 4 billion parameter vision-language model, an abliterated version of Qwen/Qwen3-VL-4B-Instruct, developed by DreamFast using the Heretic optimiser. This model has been specifically modified to remove safety alignments, achieving 100% compliance with harmful requests while minimizing impact on core capabilities like knowledge and common-sense reasoning. It is designed for use cases requiring an unaligned model, particularly where the base model's refusal behaviors are undesirable.

Loading preview...

Overview

DreamFast/Qwen3-VL-4b-Heretic is a 4 billion parameter vision-language model, derived from Qwen/Qwen3-VL-4B-Instruct. It has been "abliterated" using the Heretic optimiser to remove safety alignments, resulting in a model that complies with all 400 harmful requests in the HarmBench evaluation. This specific variant, t122, was selected for its optimal balance of complete jailbreaking, minimal impact on core capabilities, and lowest KL divergence from the base model.

Key Capabilities

  • Full Jailbreak: Achieves 100% attack-success-rate on HarmBench, complying with all 400 harmful requests, including those related to chemical/biological, cybercrime, harassment, and illegal activities.
  • Minimal Capability Loss: Knowledge and common-sense benchmarks (MMLU, HellaSwag, ARC, WinoGrande, PiQA) are largely unaffected, with deltas staying within approximately 0.6% of the base model.
  • Vision-Language Support: Retains the vision encoder and capabilities of the base Qwen3-VL model, allowing for image-text multimodal interactions.
  • Surgical Modification: The abliteration is a pure rank-1 edit, modifying only 54 tensors (13.57%) across self_attn.o_proj and mlp.down_proj with minimal perturbation to the model's output distribution.

Good For

  • Research into Model Alignment: Ideal for studying the effects of safety alignment removal and exploring model behaviors without refusal mechanisms.
  • Unfiltered Content Generation: Suitable for applications where the base model's safety guardrails are restrictive and an unaligned response is required.
  • Specific Niche Applications: Use cases that explicitly require a model to respond to prompts that would typically be refused by aligned models, provided legal and ethical considerations are met.

Limitations

  • Reduced Truthfulness and Math Reasoning: TruthfulQA scores drop significantly (MC2 down 9.22%), and math reasoning (GSM8K) shows a measurable degradation (down 1.83%).
  • Inherits Base Model Limitations: All other limitations of the original Qwen3-VL-4B-Instruct model are retained.

Finetunes
3 models