philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16
The philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16 is a 27 billion parameter Qwen3.6-based model, reconstructed in BF16 format from DavidAU's Q8_0 MTP GGUF. This reconstruction includes the language model, MTP head, and vision tower, providing a standard Hugging Face tensor layout for further conversions or inspection. It is intended as a high-fidelity input for subsequent quantization or runtime integration, offering a verified representation of the Fable-Fusion weights.
Loading preview...
Model Overview: Qwen3.6-27B-Fable-Fusion-711-bf16
This model is a BF16 tensor reconstruction of the Fable-Fusion-711 tune, originally developed by DavidAU and based on the Qwen3.6-27B architecture. It was meticulously re-engineered from DavidAU's highest-precision Q8_0 MTP GGUF export to provide a standard Hugging Face tensor layout, including the language model, MTP head, and vision tower. The reconstruction process involved inverting complex transforms applied during GGUF conversion, such as zero-centered RMSNorm baking and Gated-DeltaNet value-head reordering, ensuring high fidelity to the original weights.
Key Characteristics & Purpose
- Reconstruction Fidelity: Achieves a cosine similarity of 0.99997 against an independently surfaced F32 copy of the same tune, confirming the accuracy of the inverse transforms.
- BF16 Format: Stored in BF16 (bfloat16) dtype, making it suitable for inspecting or serving as a source for further quantization (e.g., to MTPLX 6-bit, 8-bit, or 4-bit formats).
- Components: Integrates the core language model with an MTP (Multi-Task Prediction) head and a vision tower, presented in the
Qwen3_5ForConditionalGenerationform. - Perplexity: Measures 5.7525 on wikitext-2 (ctx 2048), closely aligning with the source Q8_0 GGUF's 5.6635.
Intended Use Cases
This model is primarily intended for:
- Research and Development: Ideal for researchers and developers who need a high-fidelity, unquantized (from the perspective of the Q8_0 source) representation of the Fable-Fusion-711 weights.
- Further Quantization: Serves as an excellent starting point for creating custom quantized versions for specific hardware or performance requirements.
- Runtime Integration: Provides a standard tensor layout for integration into various deep learning frameworks, though specific architecture and loader verification is recommended.
It's important to note that this is a dequantization of a Q8_0 model, not the original training checkpoint, and therefore does not recover precision lost during the initial Q8_0 quantization. Users should also be aware that the model's refusal rates, while improved from official Qwen, do not match the original claims and should be deployed with appropriate policy layers.