NVIDIA ACE Game Agent SDK: Unreal Engine 5 plugins, DLSS 4.5

NVIDIA announced expanded RTX integration for Unreal Engine 5 at Unreal Fest 2026, releasing the NVIDIA ACE Game Agent SDK Beta and a suite of Unreal Engine 5 plugins for automatic speech recognition, small language models, and text-to-speech, plus a new DLSS 4.5 Unreal Engine plugin. The company published the announcement on June 16, 2026 and is positioning these tools for on-device, low-latency AI companions and NPCs.

What is the NVIDIA ACE Game Agent SDK?

The NVIDIA ACE Game Agent SDK is a lightweight, open source C/C++ agentic framework designed for native in-game integration and optimized for small models on NVIDIA RTX hardware. Developers interact with three core API categories: an Agent API that owns chat history and drives multistep, tool-assisted reasoning; a Chat API for stateless inference control; and a RAG API for semantic, lexical, and hybrid knowledge retrieval to ground responses in developer-built databases.

The SDK has been battle-tested in real games. Examples cited include PUBG: BATTLEGROUNDS, where KRAFTON’s Ally runs fully on-device on GeForce RTX and will enter open beta, and Total War: PHARAOH, whose experimental in-game advisor uses a RAG architecture that queries 1,200+ interlinked game data tables in real time and whose playtest program launches in 2026.

What do the new ACE Unreal Engine 5 plugins include?

The new ACE Unreal Engine 5 plugins provide local, runtime AI models with both Blueprint and C++ support that cover automatic speech recognition, small language models, and text-to-speech. The ASR plugin includes a ready-to-use English model, nemo-conformer-ctc-120m, with download options for seven additional languages and Blueprint examples. The small language model support offers local GGUF support, low-latency text generation, function calling, and includes a ready-to-use Qwen 3.5 4B model. The TTS plugin ships with the Chatterbox Turbo 350M model, example voices, and sample levels.

NVIDIA positions these plugins as local, RTX-optimized workflows to avoid the high latency and variable operational cost of cloud services. The company also released a DLSS 4.5 Unreal Engine plugin that adds Dynamic Multi Frame Generation, Multi Frame Generation 6X Mode, and the second-resolution transformer for Super Resolution, with updated APIs and sample code.

How are other tools and studios using this stack?

NVIDIA highlighted additional integrations and demos. Animotive Kimodo is an open source project for promptable, controllable human motion; an Animotive Kimodo Unreal Engine plugin ships today, with Blender and Maya integrations to follow. SOKRISPYMEDIA is building a full-scale first-person Chalk Warfare game that uses local AI for on-the-fly weapon creation, and trained bespoke models on a Puget Systems workstation built around dual NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs. NVIDIA also tied ACE to workflows that use synthetic training data and LoRA-powered performances for MetaHuman NPCs at Unreal Fest sessions.

Why it matters

Moving ASR, SLMs, TTS, retrieval, and agent logic on-device reduces conversational latency and keeps inference tightly coupled to live game state, which matters for responsive NPCs like PUBG’s Ally and the Total War advisor. Local, RTX-optimized plugins also let studios avoid ongoing cloud costs and unpredictable latency while shipping a complete NPC pipeline into Unreal Engine 5 projects.

What to watch

Register for NVIDIA’s Level Up webinar, Building AI-Powered NPCs in Unreal Engine 5.7, on June 30 for more plugin details and live Q&A, and watch the Total War: PHARAOH playtest program launching in 2026 for a public test of RAG-driven in-game advisors. NVIDIA AI research leaders Sanja Fidler, Ming-Yu Liu, and David Luebke will present related graphics and simulation work at SIGGRAPH 2026 on July 20.