LLMUnity vs Neocortex - Local LLM NPCs in Unity

Quick answer: LLMUnity is a free library that runs open-source LLMs inside a Unity build, so text generation happens entirely on the player's device. If you need voiced characters, a character editor your writers can use, or more than one engine, Neocortex covers those and still lets you prototype with local models in Unity.

If you have searched for "run an LLM locally in Unity", you have found LLMUnity. It is an open-source package from undream.ai, licensed under Apache-2.0, that runs models through llama.cpp inside your game with no server and no data leaving the device. This post covers what it gives you, the trade-offs of local and cloud inference, and when an LLMUnity alternative saves your team time.

LLMUnity in one minute

• Runs on PC, Android, iOS and VR, tested on Unity 2021 LTS through Unity 6.
• Hardware: CPU inference, plus GPU support for NVIDIA, AMD and Apple Metal.
• Models: a built-in manager downloads or loads GGUF models, with LoRA support. A "Download on Build" option fetches the model on first launch instead of bundling it.
• Retrieval: a RAG system with approximate nearest neighbour search, and embeddings from the model.
• Structured output: grammars and JSON schema to constrain replies, plus a function-calling sample.
• Chat history: save and load conversations.
• Remote mode: you can run a server build and have characters in another project connect to it by host address.
• Price: free for personal and commercial use.

Can LLMUnity run without a server?

Yes, and that is the point of it. Everything runs inside the build, which means no API bill and no network dependency.

It also means every decision a hosted platform makes for you lands on your team. You choose the model, test it on your lowest-spec target, and keep it inside the memory left over after your game's own assets. On phones, LLMUnity's guidance is tiny models of about 1 to 2 billion parameters, which limits how well a character holds a persona over a long conversation. You write the prompts, the memory logic and the guardrails. There is no text-to-speech or speech-to-text pipeline in the package, so voice is a separate integration. Characters are configured in the Inspector and in C#, so a writer who wants to change a personality needs a programmer or an editor build. And if a player's device cannot run the model, the fallback is a server you host yourself using the remote mode.

Local, hybrid or cloud: picking an approach

The middle column is the route Neocortex supports in Unity. You can run local LLMs in Unity through Ollama for text chat while you build, then point the same project at our cloud for players. If your studio needs the whole platform on its own servers, self-hosting is available as a Custom enterprise arrangement.

One NPC, two routes

Take a merchant in a PC survival game who haggles, remembers what she sells, and warns the player about a storm.

With LLMUnity, you pick a GGUF model that fits next to your game in memory, set up the RAG system with her price list and lore, and write a system prompt in C#. For haggling, you constrain her replies with a JSON schema so your code can read the offered price. For a voice, you add a separate TTS package and a speech recognition package, and handle the latency between all three. Each change to her personality goes through a programmer.

With Neocortex, a writer builds her in the browser: Personality for her tone, Knowledge with the price list as a PDF, Intents for buying and selling, and an Action such as OFFER_PRICE your game reads. Your game sends the storm as a recent event. She answers with a line, an emotion tag and her action, in a streamed voice you designed from a text description. While you build, her text chat can run on a local model through Ollama; at launch she runs in our cloud with her voice.

LLMUnity vs Neocortex

Neocortex also handles the parts of a game conversation that sit above the model. Conversation Flows branch a scene into story beats and return checkpoint keywords (flowState) your game reacts to. Group chat puts up to six characters in one conversation on Pro and Team. Unity WebGL builds get microphone input, so browser players can speak to characters too.

Where to go from here

If your game must run with no network at all, LLMUnity gives you inference and you build the character layer around it. If you want characters that talk, act and sound right without building that layer, start in the Neocortex editor: build one character, test it in the browser for free, and run it in a Unity sample scene with a local model or our cloud behind it.

Details about LLMUnity were last checked in October 2026.

Frequently Asked Questions