Summary
GET /v1/models currently only returns the single model loaded via -m. It should also list models available in the MACAFM_MLX_MODEL_CACHE directory so clients can discover what's available without loading each model first.
Motivation
When integrating AFM with benchmark tools (e.g. ToolCall-15), clients need to discover available models programmatically. Currently, the only way to know what models are cached locally is to scan the filesystem directly.
Proposed behavior
- Scan
MACAFM_MLX_MODEL_CACHE for models--<org>--<name>/ subdirectories
- Return them in the
/v1/models response alongside the currently loaded model
- Distinguish loaded vs available models (e.g.
loaded: true vs loaded: false)
- The loaded model should still appear first in the list
Context
Discovered while integrating AFM as a provider in ToolCall-15, a tool-calling benchmark dashboard. Model discovery would enable auto-populating the model selection UI.
Summary
GET /v1/modelscurrently only returns the single model loaded via-m. It should also list models available in theMACAFM_MLX_MODEL_CACHEdirectory so clients can discover what's available without loading each model first.Motivation
When integrating AFM with benchmark tools (e.g. ToolCall-15), clients need to discover available models programmatically. Currently, the only way to know what models are cached locally is to scan the filesystem directly.
Proposed behavior
MACAFM_MLX_MODEL_CACHEformodels--<org>--<name>/subdirectories/v1/modelsresponse alongside the currently loaded modelloaded: truevsloaded: false)Context
Discovered while integrating AFM as a provider in ToolCall-15, a tool-calling benchmark dashboard. Model discovery would enable auto-populating the model selection UI.