Model catalog

What it does

The Model Catalog shows all available models searchable by family (Qwen, Llama, Mistral, etc.), modality (text, vision, embedding), parameters, and tool-calling support. You can view model details (context window, quantization format, etc.) and load models into the serving layer. The catalog federates from multiple sources: LocalCatalog (static seed), Hugging Face, ModelScope, and Ollama Hub.

Prerequisites

• A running stack with access to at least one federation source (Ollama Hub is built-in; HF/ModelScope require API keys in .env).

Steps

1. Open the Catalog page

Click Catalog in the left rail (top of the icon grid). The page displays search filters on the left and a table of models on the right.

2. Search and filter models

Use the filters to narrow down:

• Family — Qwen, Llama, Mistral, GPT, etc. (multi-select).
• Modality — Text, Vision, Embedding, Audio (multi-select).
• Parameters — slider to pick a range (e.g. 3B-7B).
• Tool calling — toggle to show only models that support function calls.
• Format — safetensors, GGUF, AWQ, GPTQ, Ollama (multi-select).

Click Search (or the filters auto-apply). Results update in the table.

3. View model details

Click a model row to open a side drawer with:

• Model id — full identifier (e.g. qwen2.5:3b).
• Description — source and brief summary.
• Parameters — total model size (3B, 7B, etc.).
• Context window — max input tokens supported.
• Quantization format — GGUF, AWQ, etc.
• Tool calling — yes/no.
• Federation source — where it came from (LocalCatalog, HF, Ollama, etc.).

4. Load a model

In the drawer, click Load. The model is downloaded (if not already cached) and registered in the serving layer. A toast confirms the load.

5. Check loaded models

Once loaded, the model appears in:

• The model dropdown in the Chat panel.
• GET /v1/models API endpoint.
• The By Model tab in the Cost dashboard.

(Direct API) Search the catalog

curl -s -H "Authorization: Bearer $BIGWAVE_TOKEN" \
  "http://localhost:8080/admin/catalog/search?family=qwen&modality=text&min_params=3&max_params=7" | jq '.models'

Returns a list of models matching the filters.

(Direct API) Get model details

curl -s -H "Authorization: Bearer $BIGWAVE_TOKEN" \
  "http://localhost:8080/admin/catalog/models/qwen2.5%3A3b" | jq

Returns full model info including revisions and quantization options.

Verify

• Open the Catalog page; the search results load within ~1 second.
• Search for "Qwen"; results narrow to Qwen models.
• Click a model; the detail drawer shows full metadata.
• Click Load; a toast confirms the load.
• Open the Chat panel; the model appears in the model dropdown.
• Send a chat with the new model; the inference is logged in Observability.

Common issues

Symptom	Cause	Fix
Search returns 0 results	Filters are too restrictive or no federation sources are configured	Loosen filters; check .env for HF_TOKEN and MODELSCOPE_TOKEN if using those sources
"Model not found" on load	Model id is invalid or doesn't exist in any federation source	Try a different model from the search results; verify the model exists on Hugging Face or Ollama Hub
Model loads but chat fails with "Model not found"	Model finished loading but serving layer hasn't registered it	Refresh the Chat panel dropdown; if still missing, check api container logs
Search is slow (>5s)	Federation sources are slow or down (network issue)	Retry the search; check internet connectivity and federation service health
Ollama Hub models are missing	Ollama server is not running or is on a different host	Check that the local Ollama instance is accessible; curl http://localhost:11434/api/tags should list models

Related

• Send a chat message — requires a loaded model to chat with.
• Overview dashboard — the Catalog node shows available model count.
• Tool calls inspector — filter to models that support tool calling using the Catalog filter.