Features:

On-device LLM runtime using GGUF models. No cloud, no API keys.
Blueprint-callable Engine Subsystem (UMagiLLMSubsystem): LoadModel, Say, SetPersona, ClearConversation, IsReady, OnToken, OnDone.
Streaming token output for real-time NPC dialogue and chat UI.
Background worker thread; inference never blocks the game thread.
10 preset language personas (en/ko/ja/zh/de/fr/es/it/pt/ru); add more via Persona.<tag>.txt in Config/MagiLLM/.
Bundled Gemma 4 GGUFs (E4B Q4_K_M ~4.7 GB default; E2B Q5_K_M ~3.2 GB); custom GGUFs supported.

Code Modules:

Number of Blueprints: 0 Number of C++ Classes: 3 Network Replicated: No

Supported Development Platforms:

Supported Target Build Platforms: Win64

Important/Additional Notes:

Built for Unreal Engine 5.7.
Local GGUF via llama.cpp; no OpenAI/Anthropic/Gemini key required.
Third-party: llama.cpp (MIT) and Gemma 4 (Gemma Terms of Use). Both licenses included.
Bundle adds ~7.8 GB and needs comparable RAM at runtime.
CPU-only on Win64.

MagiForge

Product types