Qwen3.5-4B fine-tune wows with eerie prose

More proof that 4B-class open models are now good enough to be shaped into surprisingly capable local specialists. The sample is anecdotal, but it shows the bottleneck is shifting from “can the model do it?” to “can you tune and run it well?” Qwen3.5-4B is an official open-weight target, so it makes sense for hobbyist SFT/LoRA experiments. The interesting signal here is coherence and scene control, not just prettier wording. Local latency is still the pain point; inference stack and hardware matter as much as the fine-tune. This is most useful for narrow writing tasks, domain assistants, or on-device prototypes. Treat it as a qualitative demo, not a benchmark for general capability.