Local and On-Premise AI NPC Platforms for Studios

Local and on-premise AI NPC platforms come in two kinds: on-device SDKs that run the model on the player's hardware, and on-premise platforms that run the character backend on servers you control. This roundup covers nine options across both, with a decision guide for cloud, on-premise and on-device deployment.

The reasons vary. A training simulator for a secure site cannot send voice to a third-party cloud. A museum kiosk sits on an unreliable network. A publisher wants the backend inside its own infrastructure. A team that watched AI character vendors shut down wants a plan that does not depend on someone else's servers.

On-premise AI NPC platforms at a glance

The nine options

Tryll Engine

Tryll Engine runs language models, speech recognition, text-to-speech, vision models and agents on the player's GPU, detecting NVIDIA, AMD and Intel hardware. Unreal and Unity are listed as active with alpha access on request, and a native C++ route is planned. Features include knowledge from game docs and lore, persistent NPC memory, agent actions and runtime content generation. It picks a model between 4B and 14B parameters with quantization and needs 8GB of VRAM, with 12GB or more recommended. Pricing is free to start, then about $0.08 per copy sold at a $15 price, scaling with the price. It is backed by Epic MegaGrants and Early Game Ventures, and no shipped games are listed. More in Tryll Engine vs Neocortex.

NVIDIA ACE Game Agent SDK

The Game Agent SDK is an open-source C and C++ framework, in beta, for in-game agents running small models on RTX PCs. It exposes an Agent API with chat history and tool use, a stateless Chat API, and a RAG API for game data. The Unreal Engine 5 plugins add local speech recognition, GGUF language models with function calling (Qwen 3.5 4B is included), and Chatterbox Turbo text-to-speech. No Unity plugin is listed. See our NVIDIA ACE alternative for Unity.

Snapdragon Game AI SDK

Qualcomm's first dedicated gaming AI SDK, announced at GDC, runs speech recognition, quantized language models and text-to-speech on the Snapdragon NPU, CPU and GPU. It targets Android phones and Windows on Snapdragon PCs, with XR devices listed as coming, and Qualcomm frames it around NPCs, AI teammates, companions and coaches. Features currently ship as Unreal Engine 5 plugins for 5.6 and 5.7. The plugin source is BSD 3-Clause, while SDK binaries and models carry their own terms. Players on other chips need another path. Read Snapdragon Game AI SDK vs Neocortex.

X&Immersion

X&Immersion, based in Paris, sells local AI tools: DiaGen for dialogue and in-game text, Ariel for voice and Geppetto for lip-sync and facial animation. Runtime versions run on the player's device or on private servers, and Ariel and Geppetto can also run through an API call. Its site lists Unity, Unreal Engine and Wwise integrations. Pricing is not published and starts with a demo request. See our X&Immersion alternative.

LLMUnity

LLMUnity is an open-source Unity package, Apache-2.0, that runs GGUF models through llama.cpp on PC, mobile and VR, fully offline. It includes retrieval, function calling through grammars and saved chat history, plus a remote mode that points characters at a server build you host. There is no voice, and on phones it recommends models of about 1 to 2 billion parameters. Compare it in LLMUnity vs Neocortex.

NobodyWho

NobodyWho is a Rust inference engine on llama.cpp, licensed EUPL-1.2, free for commercial use. It ships a Godot plugin plus bindings for Kotlin, Swift, React Native, Flutter and Python, and runs on desktop, Android and iOS, with no web export. It supports tool calling, Whisper speech-to-text and local text-to-speech. Our Godot AI NPC roundup covers it in context.

Convai on-premise

Convai's Enterprise plan lists isolated cloud, private and on-premise operating models, with the Convai backend and AI services running in your own data center. Pricing and SLAs are custom. Its self-hosted documentation covers the components, including an auth server with SSO, a REST API server and a WebRTC conversation service, with guides for GCP, AWS and Kubernetes. See our Convai alternative.

Meta AI Building Blocks

Meta's AI Building Blocks are Unity components in the Meta XR SDK, for Unity 6 and Quest 3 or 3S. Small language models such as SmolLM, Qwen and Phi can run on the headset, and object detection runs through Unity's Inference Engine. Speech-to-text and text-to-speech use cloud providers such as OpenAI and ElevenLabs, so a fully offline voiced character is not covered. Read Meta Quest AI Building Blocks vs Neocortex, and for Unity's own runtime, Unity Sentis vs Neocortex.

Neocortex

Neocortex is an AI character platform. The self-serve product runs in our cloud with a visual node editor, voice and SDKs for Unity, Unreal Engine 5 and Godot (beta), plus a Web API. On a Custom arrangement, the same platform runs on your own servers, air-gapped, or inside your game build.

Decision guide: cloud, on-premise or on-device

Choose cloud if your game is online anyway, your players run a wide range of hardware, and you want writers editing characters today.

Choose on-premise if data cannot leave your network, the site has no reliable internet, or a client contract puts the backend in your infrastructure. Training simulators, defence, healthcare, museums and events usually land here.

Choose on-device if the character must work for players offline and you can set a hardware floor and test on each device tier.

Many projects mix them: cloud while prototyping, then on-premise or on-device once requirements are fixed. Our AI NPC pricing comparison covers how each cost shape adds up.

What running on the player's hardware costs

On-device inference moves the bill from a server to the player's machine, and the game pays for it in other ways.

• Shared memory. The model sits in the same VRAM or RAM as your textures and meshes.
• Frame budget. Inference competes with rendering and physics, so it has to be scheduled around the frame.
• Smaller models. What fits next to a game is smaller than what a server runs, and it shows in long conversations.
• Per-device testing. Every GPU tier, phone or headset needs its own pass for speed, memory and heat.
• Model licences. Each model you ship has its own terms, and speech models vary more than language models.

Measured speed. On a laptop RTX 3080 (8 GB), our on-device runtime returns its first token in 0.06 seconds and starts speaking in about one second, because it voices the first sentence while the rest is still being generated. Speech recognition and text-to-speech run on the CPU, so a 3B model needs about 3 GB of VRAM, and a 2B model generates over 300 tokens per second in under 1 GB.

Your players' hardware sets the final numbers, so test on your lowest target. Our cloud product is built for real-time conversation, and on-premise deployments are sized with each studio for the load they expect.

Where Neocortex fits

Neocortex gives studios a choice of where characters run, with the same character definition in each case. On Custom (enterprise) arrangements, set up through a conversation with our team:

• Full platform on your own servers or private cloud: the character editor, the API, your characters and Conversation Flows.
• Local voice: speech-to-text and text-to-speech run on-premise, not only the language model.
• Air-gapped operation: fully offline, for secure sites, training simulators, kiosks and events.
• On-device inference: inference can ship inside the game build and run on the player's machine.
• Custom support: our team works with you on deployment, model choice and integration.

If your project needs any of these, contact us.

Characters are built the same way in every deployment: a visual node editor with Personality, Emotion, Knowledge (PDF or TXT), World, Objects, Intents, Actions and Voice. Conversation Flows return checkpoint keywords (flowState) your game reacts to, and replies carry emotion tags and actions such as OPEN_DOOR. The self-serve product adds voice cloning and voice design, group chat with up to six characters on Pro and Team, and plans from free to Team.

Next steps for your studio

1. Write down where your characters must run: your cloud, your servers, an air-gapped site or the player's device.
2. List your target hardware and decide whether a VRAM or NPU floor is acceptable.
3. Build one character in the Neocortex editor and test it in the browser for free; in-editor testing never uses credits.
4. Run it in Unity, Unreal or Godot through our SDKs to see the integration.
5. Contact our team to plan an on-premise, offline or on-device deployment around your project.

For every head-to-head comparison in one place, see Neocortex comparisons and our earlier AI NPC tools with local AI support.

Competitor details were last checked in October 2026.

Frequently Asked Questions