YFC-112358/Qwen3.6-27B-Della-Deckard-Isometry-Geodesic-v2

Model Overview

YFC-112358/Qwen3.6-27B-Della-Deckard-Isometry-Geodesic-v2 is a 27 billion parameter Qwen3.6-based model, developed by YFC-112358, that applies a linear task-vector merge to reduce generated 'thinking' tokens. This model is a direct merge of Qwen3.6-27B-Della-Deckard-v1 with a terseness task vector from maldv/Qwen3.6-27B-Isometry-Terse, specifically targeting 400 two-dimensional projection tensors. The merge is notable for its simplicity, as the two contributing edits were found to be orthogonal, allowing for a straightforward linear addition without complex interference resolution.

Key Capabilities & Performance

• Significant Token Compression: Achieves a 7.1x reduction in 'think' tokens on GSM8K (from 539.5 to 76.0 median tokens) compared to its Della-Deckard-v1 predecessor, and 14.9x against the official Qwen/Qwen3.6-27B base model.
• Maintained Accuracy on Simple Tasks: GSM8K pass@1 accuracy did not measurably drop (0.75 to 0.85, though on a small sample size).
• Identical Vision Tower: The vision tower is byte-identical to the official Qwen3.6-27B, implying no degradation in multimodal capabilities, though no specific multimodal evaluation was performed.

Use Cases

• Good Fit:
  • High-volume short-answer workloads where models might over-generate 'thinking' tokens.
  • Latency-bound interactive applications.
  • Cost-bound batch inference, due to reduced token generation.
• Bad Fit:
  • Competition math, multi-hop reasoning, or long agentic chains where deliberation is desired, as it shows regression on harder multi-step problems.
  • Any task requiring high accuracy on complex reasoning, as the model failed acceptance gates for hard problem accuracy and think length.