Muddy Terrain Games

Products Docs Careers Support

Quick Start

Get a working AI response inside Unreal Engine in five minutes. Two paths — pick whichever matches what you already have installed.

  • HTTP path (fastest, ~2 min) — You install a local inference server (Ollama, LM Studio, etc.) and point the plugin at it. No plugin rebuild required.
  • Embedded path (~5 min) — The plugin runs llama.cpp directly inside the game process. No external server, fully offline. Requires dropping prebuilt binaries into the plugin and rebuilding.

If you’re unsure which to pick, start with HTTP. Both use the same Blueprint nodes — switching is a single dropdown change.

Path A: HTTP Provider (Ollama)

Step 1: Install Ollama

Download and install Ollama. It runs a local HTTP server at http://localhost:11434 in the background once installed.

Step 2: Pull a Model

In a terminal:

ollama pull llama3

This downloads Meta’s Llama 3 (~4.7 GB). For a faster test, use ollama pull gemma3:1b (~800 MB).

Step 3: Enable the Plugin

In Unreal, open Edit > Plugins, search for GenAI Llama, enable it, and restart the editor if prompted.

Step 4: Wire Up the Blueprint

  1. Create a new Actor Blueprint and open it.
  2. On Event BeginPlay, drag off and add Request Chat Stream.
  3. On the Chat Settings pin, right-click and Split Struct Pin, then set:
    • Model = llama3
    • Connection > Provider = Ollama
    • Connection > Base URL = http://localhost:11434
    • Messages = one entry with Role = user and Content = Tell me a joke.
  4. On the On Event delegate, add a Print String connected to Partial Message > Content.
  5. Drag the Blueprint into the level and press Play.

Tokens will stream into the top-left viewport overlay. Done.

Path B: Embedded Inference (llama.cpp)

Pinned to llama.cpp b8802 (April 2026)

Embedded inference is validated against this exact build. llama.cpp renames and removes exported symbols between releases, so a newer version will typically fail symbol resolution at startup. The Project Settings > Plugins > GenAI Llama panel has a one-click button to open the pinned release page.

Step 1: Download the Pinned llama.cpp Binaries

Open Project Settings > Plugins > GenAI Llama and click Download b8802. Pick the zip matching your platform:

Platform Archive to download Notes
Windows llama-b8802-bin-win-cpu-x64.zip CPU-only — works on any Windows machine. For NVIDIA, use the -cuda-* variant instead.
macOS (Apple Silicon) llama-b8802-bin-macos-arm64.zip Metal built in.
Linux llama-b8802-bin-ubuntu-x64-cpu-*.tar.gz CPU baseline.

Step 2: Drop the Binaries Into the Plugin

Extract all shared library files and copy them into:

<YourProject>/Plugins/GenAILlama/ThirdParty/LlamaCpp/lib/<Platform>/

Where <Platform> is Win64, Mac, or Linux.

You should end up with files like llama.dll, ggml-base.dll, ggml-cpu.dll, ggml-rpc.dll in the Win64 folder (or their .dylib / .so equivalents on Mac / Linux).

macOS only: remove the quarantine flag

Dylibs downloaded through a browser are flagged by Gatekeeper and silently refuse to load. Run once after copying:
xattr -cr <YourProject>/Plugins/GenAILlama/ThirdParty/LlamaCpp/lib/Mac

Step 3: Download a Tiny Test Model

Qwen2.5-0.5B-Instruct Q2_K is ~280 MB and runs on any machine:

  1. Go to Qwen2.5-0.5B-Instruct GGUF on Hugging Face.
  2. Download qwen2.5-0.5b-instruct-q2_k.gguf and drop it into <YourProject>/Content/Models/.

More models + quantization guidance: GGUF Models.

Step 4: Rebuild the Project

Close the editor (so binaries can update on Windows), then rebuild from your IDE or via Build in the Unreal Engine launcher. The build log should show:

GenAILlama: llama.cpp shared libraries found at <path>. Embedded inference enabled (dynamic loading).

Re-open the editor. Project Settings > Plugins > GenAI Llama should now display Embedded Inference — Loaded in green.

Step 5: Wire Up the Blueprint

  1. Create an Actor Blueprint and open it.
  2. On Event BeginPlay, add Load Embedded Model and set:
    • Model Alias = test-model
    • Model Path = Models/qwen2.5-0.5b-instruct-q2_k.gguf (relative to Content/)
    • GPU Layers = 0 (CPU only — safe on any machine)
    • Context Size = 2048
    • Thread Count = 0 (auto)
  3. On its On Complete pin, add Request Chat Stream with:
    • Connection > Provider = Embedded (llama.cpp)
    • Model = test-model (the alias from step 2)
    • Messages = one entry with Role = user and Content = Tell me a joke.
  4. On On Event, add a Print String connected to Partial Message > Content.
  5. Drop the Blueprint into the level and press Play.

The Output Log will show Loaded model: qwen2.5-0.5b-instruct-q2_k.gguf and tokens will stream to the viewport.

No NVIDIA GPU? Keep GPU Layers at 0.

The CUDA backend only loads if ggml-cuda.dll is present and the NVIDIA driver/runtime is installed. Non-zero GPU Layers on a CPU-only setup will fail the load. Use CPU first, then switch to a GPU backend later.

Next Steps

  • Blueprint Nodes — Every node the plugin adds, with inputs/outputs.
  • C++ API — Production-ready C++ snippets with lifetime-safe patterns.
  • GGUF Models — Where to find GGUFs, which quantization to pick, sizing for mobile vs desktop.
  • Embedded Inference Setup — Full per-platform setup, compile-from-source, GPU backends.
  • Troubleshooting — When something doesn’t work, start here.
× Full-size image