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 language model, reconstructed by philipjohnbasile from DavidAU's Q8_0 MTP GGUF. This bf16 full-precision version includes a language model, MTP head, and vision tower, designed for developers to create custom quants or run in transformers/vLLM. It offers a perplexity of 5.7525 on wikitext-2 and is intended for research and creative work, particularly where a high-fidelity starting point for quantization is needed.

Loading preview...

Model Overview

This model, philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16, is a 27 billion parameter Qwen3.6-based language model. It is a full-precision bf16 reconstruction of DavidAU's Fable-Fusion-711 tune, which was originally published only as a Q8_0 MTP GGUF. The reconstruction includes the language model, an MTP head, and a vision tower, presented in a standard Qwen3_5ForConditionalGeneration format.

Key Characteristics

  • Reconstructed bf16 Checkpoint: This model provides the bf16 safetensors checkpoint that was not originally published by DavidAU, derived from his highest-precision Q8_0 MTP GGUF.
  • Integrated Capabilities: Features a language model, MTP head, and vision tower, making it a multimodal foundation.
  • Performance: Achieves a perplexity of 5.7525 on wikitext-2 (context 2048), closely matching the original Q8_0 GGUF's 5.6635.
  • Provenance: While a dequantization of Q8_0, it has been independently validated against an F32 copy of the same tune, showing a cosine similarity of 0.99997 on mtp.fc.weight and 1.00000 on model.language_model.norm.weight.

Intended Use Cases

  • Custom Quantization: Ideal for developers who wish to create their own quants (e.g., MLX, EXL, AWQ, GPTQ) from a high-fidelity starting point.
  • Direct Deployment: Suitable for running the tune directly in transformers or vLLM environments without relying on GGUF conversions.
  • Research and Creative Work: The model is intended for research and creative applications, with a note that its "Heretic" nature ablates refusal behaviors, requiring users to deploy it behind their own policy and logging layers.

Finetunes
1 models