Qwen3.5 Small hits Mac throughput wall

This is a throughput problem, not a RAM problem. On Apple Silicon, extra model copies usually compete for the same unified memory and GPU time unless the serving layer batches requests. mlx-lm already points developers toward batch generation, prompt caching, and distributed inference, which suggests the right fix is workload batching rather than model duplication. The community has already asked for batch and parallel processing in mlx_lm.server, so this bottleneck is a known gap in the stack. mlx-vlm's docs focus on one-model-at-a-time serving with dynamic loading, which makes it flexible but not a high-concurrency engine. For OCR pipelines, grouping documents by prompt shape and minimizing repeated context will usually beat spinning up more standalone processes.