Vision facts
Fourteen detected VLM families
The model library detects 14 vision-language model_type families.
This is an evidence-backed family count, not a guarantee that every checkpoint or processor variant will load.
Verified
Apple Silicon unified memory
MLX arrays use the Mac's shared CPU/GPU memory system.
Unified memory reduces explicit transfers between CPU orchestration and integrated-GPU compute, but model weights, activations, and KV cache still consume finite physical memory.
Verified
Eligibility-gated continuous batching
v0.6.0 batches only eligible dense-text requests under real concurrency, with automatic serial fallback.
The tagged 4-client benchmark measured 2.5–3.2× aggregate throughput. VLM, speculative decoding, Ollama, Anthropic, and embeddings remain serial.
Verified
Vision checkpoint checklist
| Category | Question | Boundary |
|---|---|---|
| Architecture | Is the exact model_type among the detected families? | A generic MLX label is not enough |
| Processor | Does the checkpoint include compatible image preprocessing? | Processor variants can differ within a family |
| Quantization | Is this quantized checkpoint supported by its architecture? | Support is not universal across weights |
| Memory | Do weights, image tokens, activations, and cache fit? | Vision adds workload beyond text weights |
| Serving path | Does the selected API path accept vision input? | Batching is dense-text-only; VLM requests remain serial |