OptimizeLLM/Qwen3.8-27B-heretic-MTP-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OptimizeLLM/Qwen3.8-27B-heretic-MTP-BF16 is a 27.8 billion parameter Qwen3.8-27B model, abliterated using HERETIC 1.4.0, retaining its MTP heads and full vision tower. This BF16 checkpoint is designed to remove refusal tendencies while maintaining the base model's original directional lean. It is intended as a source checkpoint for further quantization and use in applications requiring engagement with any topic.

Loading preview...

OptimizeLLM/Qwen3.8-27B-heretic-MTP-BF16 Overview

This model is an abliterated version of the Qwen3.8-27B base model, featuring 27.8 billion parameters, 64 layers, and hybrid linear/full attention, with native image and video capabilities. It has been processed using HERETIC 1.4.0 with local patches, employing a two-stage slot-grouped pipeline. A key characteristic is its non-refusing, not neutral behavior, achieved through directional ablation that removes the tendency to decline prompts, while still reflecting the base model's inherent viewpoint on topics. The model maintains the original Qwen3.8-27B tokenizer, ensuring byte-identical compatibility upstream.

Key Capabilities

  • Refusal Removal: Significantly reduces refusal rates, with only one refusal out of 99 on a 100-prompt harmful-behaviors set, compared to the base model.
  • Topic Engagement: Designed to engage with any topic without declining, offering a more direct interaction experience.
  • Vision Tower Intact: Retains the full vision tower and MTP heads from the original Qwen3.8-27B, suggesting potential for multimodal applications.
  • BF16 Checkpoint: Provided as a BF16 source checkpoint, suitable for further quantization to various formats like FP8.

Good For

  • Developers seeking a Qwen3.8-27B variant with reduced refusal tendencies.
  • Applications requiring a model that will engage with a wide range of topics without declining.
  • As a base for custom quantization to optimize for specific deployment environments (e.g., FP8).
  • Research into model behavior modification and safety alignment techniques.

Finetunes
1 models