The GUI uses the AutomatosX catalog
AutomatosX publishes the curated packs used by AX Engine’s GUI catalog. These packs include model metadata and required extras for their supported paths. Start with Models → Catalog and check the support matrix for the exact pack. Supported, candidate, experimental and certified describe different evidence levels; they are not interchangeable.
Example: a third-party model through the CLI
The support matrix lists mlx-community/Qwen3.5-9B-MLX-4bit for the Qwen 3.5 9B family. After installing the CLI and its download helpers, download that exact repository to a folder you control and serve it by path. This uses the CLI; it does not add the model to the GUI’s supported catalog. Allow enough memory and disk space for the pack and its cache copy.
mkdir -p "$HOME/Models" &&
ax-engine download mlx-community/Qwen3.5-9B-MLX-4bit --dest "$HOME/Models/qwen3.5-9b" &&
ax-engine serve "$HOME/Models/qwen3.5-9b" --port 31418Example: weights already on disk
For a supported, prepared MLX folder, download with a local path validates the folder and writes model-manifest.json offline. Then serve the same path. This does not convert an unsupported architecture or a raw PyTorch checkpoint into a supported model.
ax-engine download "$HOME/Models/my-model" &&
ax-engine serve "$HOME/Models/my-model" --port 31418Check format and capabilities
A Hugging Face repository ID identifies files, not a compatibility guarantee. Native serving needs a supported architecture, a supported quantization format, tokenizer files and the required model metadata. Raw checkpoints may need conversion first. Image, audio and video support also require the matching towers in the pack. Use the support matrix and /v1/models capability metadata rather than assuming every feature works.
Leave room for context and generation
Model file size is only part of the memory cost. Context, KV cache, generation and other applications also consume memory. On a 16 GB Mac, start with one compact 4-bit model and short prompts. If admission fails, use a smaller supported pack, reduce context or stop other loaded models. Historical benchmark speeds apply only to their recorded pack and host.