huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp

VISIONPricing:Input $1.06 / Cached $0.15 / Output $2.6Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp is a 27 billion parameter Qwen3.6-based model, derived from an 'abliterated' version of Qwen/Qwen3.6-27B. This bf16 build preserves the MTP (Multi-Task Prediction) head and vision tower, enabling speculative decoding and multimodal capabilities. It is designed for applications requiring uncensored text generation combined with vision processing and efficient inference via speculative decoding.

Loading preview...

Model Overview

huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp is a 27 billion parameter model built upon the Qwen3.6 architecture. It is a bf16 build derived from an 'abliterated' version of Qwen/Qwen3.6-27B, specifically llmfan46/Qwen3.6-27B-uncensored-heretic-v2. A key feature of this model is the preservation and re-grafting of the MTP (Multi-Task Prediction) head and the vision tower from the original Qwen3.6-27B.

Key Capabilities & Features

  • Uncensored Text Generation: Inherits uncensored characteristics from its heretic-abliterated base.
  • Multimodal (Vision) Support: The model.visual.* components are preserved, allowing for image-text processing.
  • MTP Head for Speculative Decoding: The re-grafted MTP head enables efficient inference using speculative decoding algorithms like SGLang/vLLM's --speculative-algo NEXTN.
  • bf16 Precision: The model is built in bf16, offering a balance of performance and memory efficiency.

Performance Metrics

  • KL Divergence: When compared to the base Qwen3.6-27B, this checkpoint shows a mean KL divergence of 0.0425 nats, indicating a close statistical distribution.
  • Perplexity: Achieves a perplexity of 7.4619 on the wikitext-2-raw test split, slightly higher than the base Qwen3.6-27B's 7.3057.

Use Cases

This model is particularly suitable for:

  • Applications requiring uncensored language generation.
  • Multimodal tasks involving both text and image inputs.
  • Scenarios where efficient inference with speculative decoding (e.g., via vLLM) is crucial for faster response times.

Finetunes
3 models