YFC-112358/Qwen3.6-27B-Della-Deckard-v1
YFC-112358/Qwen3.6-27B-Della-Deckard-v1 is a 27 billion parameter Qwen3.6-based model, created by YFC-112358, that leverages a two-stage weight merge process. It combines two reasoning-distilled sources using a DELLA merge and integrates a cross-generational task vector from a Qwen3.5 fine-tune to enhance engineering and coding abilities. This model is specifically designed to improve reasoning and coding performance while maintaining the base model's vision capabilities and Chinese language fluency.
Loading preview...
Overview
YFC-112358/Qwen3.6-27B-Della-Deckard-v1 is a 27 billion parameter model built upon the Qwen/Qwen3.6-27B base. It employs a sophisticated two-stage weight merging technique, with all mixing coefficients derived from measured weight-space distances to ensure precision and avoid arbitrary choices.
Key Capabilities & Features
- Enhanced Reasoning: Stage 1 of the merge applies a DELLA (DARE the Extreme) technique, combining two reasoning-distilled sources (
Ravionhf/qwen3.6-27b-reasoning-distill-lora-v1andnerkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill) to improve logical thinking and problem-solving. - Improved Engineering/Coding: Stage 2 introduces a cross-generational task vector from a Qwen3.5 fine-tune (
nightmedia/Qwen3.5-27B-Engineer-Deckard-Gemini) to transplant engineering and coding abilities. This was validated by a high cosine similarity (0.9478) between Qwen3.5 and Qwen3.6 base weights, indicating directional coherence. - Vision Capability: The model retains the full vision tower of the base Qwen3.6-27B model, as it was found to be bit-identical across all donor models and thus copied verbatim.
- Robust Merge Process: Utilizes
magprunefor magnitude-based pruning and sign election to prevent independently trained deltas from canceling each other out. - Chinese Language Fluency: Smoke tests confirmed fluent output for Chinese-only prompts, with no garbled tokens or code-switching, despite
embed_tokensparticipating in the merge.
Usage Considerations
- No Quantitative Benchmarks: While structural verification and a 13-prompt smoke test were performed, no quantitative benchmarks have been run. Users should treat capability claims as unproven until further testing.
- BF16 Weights: The model uses BF16 weights, totaling 51.7 GB, requiring an 80 GB GPU for unquantized use or approximately 24 GB at 4-bit quantization.
- Chat Template: Inherits the base chat template, including the
<think>block, which can be split on</think>to separate reasoning from the final answer. - Limitations: Long-context behavior beyond 16K tokens and multimodal vision capabilities were not extensively tested.