Model Catalog & Serving

The Model Library holds three kinds of model in one catalog: reference models that ship with the app, models promoted out of your training runs, and weights you upload yourself.

Custom model upload

Bring your own weights into the same catalog as everything else:

• A Keras or PyTorch checkpoint
• A GGUF file
• A LoRA adapter

Uploaded models are selectable for Testing and Inference exactly like promoted ones.

Artifact families

Family	What it is	Servable
llm_adapter	A LoRA adapter — a tuned delta, not a standalone model	No — export first
llm_hf	A full Hugging Face checkpoint	No — export first
llm_gguf	A quantized, self-contained model	Yes
Keras / YOLO weights	Vision and NLP artifacts	Used by Inference directly

Each family carries a plain-language note in the catalog, so the next action is obvious from the card.

Export and re-quantization

Export actions live on the model detail page:

• Adapter → GGUF — merge and quantize in one step, producing a servable model.
• Adapter → merged 16-bit — produce a full checkpoint you can download and use elsewhere.
• Checkpoint → GGUF — quantize a full checkpoint for the chat runtime.
• GGUF → smaller GGUF — re-quantize an existing model down a size class.

Serving

Serving starts a local runtime around a GGUF model and reports its uptime. Start, stop, and switch from the chat surface's left rail, or from the model card directly. Nothing about serving requires an account or a network connection.

Architecture Studio

For models you want to design rather than select, Architecture Studio is a visual graph editor with Keras and PyTorch emitters — build the graph, emit the code, train it like any other base.