DavidAU's Qwen 3.6 27B fine-tune achieves 100% agent success rate
Structured reasoning reduces task failures from 20% to 0% in scheduling agent.
DavidAU's latest fine-tune of Qwen 3.6 27B, developed in collaboration with other community members, delivers surprising gains for AI agents. Unlike previous solo fine-tunes, this model (often referred to as the 'Fable' variant) includes formal benchmarks showing consistent improvement over stock across multiple agentic tasks without any regression. Early adopters report that the model's reasoning block is notably more structured, often generating a clear plan before responding, which translates directly to fewer tool call errors in complex, real-world workflows.
In one detailed test, a user ran their Hermes agent 10 times on a demanding overtime scheduling task that requires logging into a complex web app with JavaScript traps. The stock Qwen 3.6 27B model failed 2 out of 10 attempts, but DavidAU's fine-tune achieved a perfect 10/10 success rate with zero failed tool calls—even at Q4 KV cache quantization on a single RTX 3090. This suggests the fine-tune is not only more reliable but also more efficient in long-context, action-driven environments. The model is available on Hugging Face and works with Unsloth's vision mmproj for multimodal needs.
- Collaborative fine-tune of Qwen 3.6 27B shows no benchmark regression and measurable agentic gains.
- Real-world test: 100% task success vs 80% for stock on a complex web scheduling agent with 262K context.
- Model produces structured reasoning plans, reducing tool call errors even at Q4 quantization.
Why It Matters
Demonstrates that well-tuned 27B models can outperform larger ones for specialized agent workloads.