philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16 is a 27 billion parameter Qwen3.6-based model, reconstructed in bf16 full-precision from DavidAU's Q8_0 MTP GGUF. This model includes the language model, MTP head, and vision tower, making it suitable for multimodal applications. It offers a high-fidelity starting point for custom quantizations and direct use in transformers/vLLM, with a perplexity of 5.7525 on wikitext-2.

Loading preview...

Model Overview

This model, philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16, is a 27 billion parameter Qwen3.6-based model. It is a bf16 full-precision reconstruction of DavidAU's Fable-Fusion-711 tune, originally released as a Q8_0 MTP GGUF. The reconstruction faithfully reproduces the language model, MTP head, and vision tower in a standard Qwen3_5ForConditionalGeneration format, making it a multimodal capable model.

Key Features and Capabilities

  • Full-Precision Reconstruction: Provides a bf16 version of the Fable-Fusion-711 tune, ideal for developers needing higher precision than the original Q8_0 GGUF.
  • Multimodal Architecture: Includes a vision tower and MTP head alongside the language model, enabling potential for combined text and image processing tasks.
  • High Fidelity: Verified against an independent F32 copy of the same tune, achieving cosine similarity of 0.99997 on key weights, confirming the accuracy of the reconstruction process.
  • Performance: Achieves a perplexity of 5.7525 on wikitext-2 (context 2048), closely matching the original GGUF's performance.

Intended Use Cases

  • Custom Quantization: Serves as an excellent starting point for users who wish to create their own MLX, EXL, AWQ, or GPTQ quants.
  • Direct Deployment: Suitable for running the Fable-Fusion-711 tune directly in transformers or vLLM environments without GGUF conversion.
  • Research and Creative Work: Designed for general research and creative applications, though users should implement their own policy layers as the model is not guaranteed to be refusal-free.

Finetunes
1 models