Local LLM Server Manager

Appearance

Quickstart Guide ​

This tutorial guides you through downloading your first language model and sending your first prompt.

Prerequisites ​

Verify these prerequisites before you begin:

  • Local LLM Server Manager is running and shows a green health indicator.
  • The Ollama engine is installed and active on port 11434.
  • Your system has at least 8 GB of available system memory or GPU VRAM.

Step 1: Open the Model Download Tab ​

Open the model search interface:

  1. Open the application window or navigate to http://localhost:5246 in your web browser.
  2. Click the Find & Download Models tab on the navigation bar.

Step 2: Download Your First Model ​

Choose one of two methods to download a language model.

Option A: Download from the Ollama Library (Fastest) ​

  1. Click the Ollama Library sub-tab.
  2. Choose one of the recommended starter models:
    • llama3.2:latest (2.0 GB): Lightweight model for general conversation.
    • qwen2.5-coder:7b (4.7 GB): High-performance model for coding tasks.
  3. Click the Pull button on your chosen model card.
  4. Observe the progress bar as the system streams the model layers to disk.

TIP

The qwen2.5-coder:7b model delivers high code generation accuracy while using under 6 GB of VRAM.

Option B: Download from Hugging Face Hub (GGUF) ​

  1. Click the Hugging Face Hub (GGUF) sub-tab.
  2. Type a repository name in the search box (for example: Qwen/Qwen2.5-Coder-7B-Instruct-GGUF).
  3. Press Enter to search the Hugging Face Hub.
  4. Select your preferred quantization file from the repository file list (for example: qwen2.5-coder-7b-instruct-q4_k_m.gguf).
  5. Click Pull Selected.
  6. Wait for the download to finish.

NOTE

Q4_K_M quantization provides an optimal balance between memory usage and generation quality.

Step 3: Inspect Telemetry and Context Size ​

Inspect your installed model and calculate memory requirements:

  1. Click the My Models tab on the top navigation bar.
  2. Locate your downloaded model in the model collection.
  3. Review the model badges (for example: Coding, Chat, 4.7 GB).
  4. Drag the Interactive KV Cache Calculator slider to your target token length (for example: 8192 tokens).
  5. Review the estimated total VRAM calculation to ensure your GPU has sufficient capacity.

Step 4: Test Your Installed Model ​

Test your newly downloaded model with an interactive prompt or automated test flight.

Option A: Verify with Real Engine Test Flight (Fastest) ​

Verify that your installed model executes properly and clears GPU memory:

  1. Click the Studio tab on the navigation bar.
  2. Locate the Test Flight panel in the studio header.
  3. Select Text from the Modality dropdown.
  4. Select or type a starter prompt.
  5. Click 🚀 Launch Test Flight.
  6. The test runner sends a verified prompt to Ollama, confirms VRAM allocation, and returns the response in seconds.

TIP

Read the complete Real Engine Test Flight Guide to test Image, Video, and Audio engines.

Option B: Send a Prompt via the REST API Proxy ​

Send a prompt from your terminal using the unified OpenAI-compatible endpoint on port 5246:

bash
curl http://localhost:5246/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5-coder:7b",
    "messages": [
      {
        "role": "user",
        "content": "Explain binary search in three sentences."
      }
    ],
    "stream": false
  }'

The server routes the request to Ollama on port 11434 and returns the generated text.

Option C: Connect External Chat Frontends ​

You can connect popular web chat frontends to Local LLM Server Manager:

  • Open WebUI: Set the Ollama URL to http://localhost:5246 or http://localhost:11434.
  • LibreChat: Add an OpenAI-compatible custom endpoint pointing to http://localhost:5246/v1.

NOTE

The built-in AI Assistant tab connects to an external gateway (such as LiteLLM or Vertex AI Gemini Flash). The assistant intentionally excludes local models from its selector to keep 100% of your local GPU memory available for heavy diffusion and creative tasks. See the AI Chat Assistant Guide to configure external credentials.

Step 5: Unload Models to Free VRAM ​

Release GPU memory when you complete your tasks:

  1. Look at the top status bar in the application.
  2. Click Unload All VRAM.

The VRAM Orchestrator sends a release signal to Ollama. The GPU memory allocation bar drops back to zero.

Next Steps ​

Released under the MIT License.

Copyright © Spelech