Benchmark data shows Qwen3.5-9B beats frontier models on OCR and field extraction, yet stumbles badly on tables. Here's the honest breakdown.