Production Case Study: Qwen 3.5 VLM on MLX for Healthcare AI

[asq-sheriff]

We deployed Qwen 3.5-4B on Apple Silicon for elderly care AI using mlx-vlm.
Key findings:

• mlx-vlm required (not mlx-lm) due to VLM architecture
• 3x latency improvement over llama.cpp for DeltaNet (20.7s → 6.9s)
• Serial queuing outperformed continuous batching for our conversational workload
• Patched chat template to disable thinking mode by default

This feedback may help with future MLX optimizations for DeltaNet architectures.
Happy to share more details about our deployment.

[Medium Link]