👉 Try the demo: https://github.com/LM-Kit/lm-kit-net-samples/tree/main/console_net/local-inference/model-catalog/host_aware_model_picker

Programmatic Model Catalog for C# .NET Applications

🎯 Purpose of the Demo

ModelCard.GetPredefinedModelCards() exposes the full set of models LM-Kit knows how to download and load. This demo turns that into an interactive picker that filters by capability (chat, vision, OCR, embeddings, speech-to-text, reasoning), sorts by file size, and then loads the chosen model with LM.LoadFromModelID().

👥 Who Should Use This Demo

A developer who wants to ship an app that adapts to the user's hardware ("smallest model that supports vision on this host").
A platform team building an admin page that shows which models are currently available.
Anyone shipping a multi-tenant product where users pick the model from a UI.

🚀 What Problem It Solves

Hard-coded model IDs do not scale. A new model is released every few weeks, VRAM budgets vary across customers, and your app needs a single source of truth for "which models can I load right now". The catalog is that source.

💻 Demo Application Overview

A console app that:

Asks for a capability filter.
Lists every catalog entry that matches, sorted by file size.
Prints ModelID, name, size, context length, quantization, capability flags.
Loads the selected model and prints architecture, parameter count, embedding size, and the cached file path.

✨ Key Features

ModelCard.GetPredefinedModelCards() for runtime catalog access.
ModelCard.Capabilities flags: Chat, Vision, OCR, TextEmbeddings, ImageEmbeddings, SpeechToText, Reasoning, ToolsCall, TextReranking, Translation.
LM.LoadFromModelID(modelID, downloadingProgress, loadingProgress) for one-line download + load.
Reports LM.Architecture, LM.ParameterCount, LM.EmbeddingSize, LM.ContextLength, LM.LayerCount, LM.GpuLayerCount, LM.HasVision, LM.HasToolCalls, LM.HasSpeechToText, LM.HasReasoning, LM.IsEmbeddingModel, LM.ModelPath.

Example Output

=== LM-Kit Model Catalog Browser ===

Pick a capability to list models for:
  1 - Chat / text generation
  2 - Vision
  ...
> 1

Found 24 model(s) matching the filter.

  # | Model ID           | Name                            |      Size | Context | Quant | Capabilities
------------------------------------------------------------------------------------------------
  1 | gemma3:270m        | Google Gemma 3 270M             |   245.3 MB |  32,768 |   4.0 | chat
  2 | gemma3:1b          | Google Gemma 3 1B                |   789.5 MB |  32,768 |   4.0 | chat tools
  ...

🏗️ Architecture

User -> capability filter
      |
      v
ModelCard.GetPredefinedModelCards()  --filter by Capabilities--> sorted list
      |
      v
LM.LoadFromModelID(selected.ModelID)  -- download + cache + open --
      |
      v
LM (ready for inference)

⚙️ Getting Started

Prerequisites: .NET 8 SDK. The first selected model downloads on first run.

Run:

git clone https://github.com/LM-Kit/lm-kit-net-samples.git
cd lm-kit-net-samples/console_net/local-inference/model-catalog/host_aware_model_picker
dotnet run

🔧 Troubleshooting

"No models match" -> your capability filter is too narrow. Pick option 7 to see everything.
Download fails -> the model is hosted on Hugging Face. Verify the host has outbound HTTPS to huggingface.co.

🚀 Extend the Demo

Add a VRAM filter that uses GpuDeviceInfo.Devices[0].FreeMemorySize to hide models that will not fit.
Pin the user's choice to a config file so the next launch skips the picker.
Add an "update" command that calls LM.LoadFromModelID() on every cached model to refresh.

Table of Contents