Unity Sentis vs Neocortex for AI NPCs

Quick answer: Unity Sentis runs neural networks inside your game. It is a building block for AI NPCs, not a character system. If you want a talking NPC this week, Neocortex gives you the character, the voice and the Unity SDK. If you need custom on-device models, Sentis is the tool, and the two work together.

Search for "AI in Unity" and Sentis comes up quickly, so it is natural to ask whether it is all you need for AI NPCs. Here is what Sentis does, what building a character on it involves, and where Neocortex fits.

Unity Sentis in brief

Sentis is Unity's library for running trained models on the player's device, on the CPU or GPU, across all Unity runtime platforms.

• Imports ONNX models, LiteRT (formerly TensorFlow Lite) models, and PyTorch exported programs.
• Tokenizes text for language models, with a tokenizer that reads the tokenizer.json files found in most Hugging Face model repositories.
• Has had two names. For a few releases the package was shown as Inference Engine. From version 2.4 the display name is Sentis again, and the package ID is still com.unity.ai.inference.

Sentis is meant for custom on-device models: vision, pose estimation, classifiers and small generative models. Unity has no conversational character feature built on it. Unity AI, the successor to Muse, is an Editor assistant for building your game, not a runtime character system.

What a talking NPC on Sentis takes

Here is the work, in order, for one NPC the player can speak to.

1. Pick a language model small enough to run at frame rate on your weakest target device, and good enough to stay in character.
2. Convert and validate it. Export to a supported format, check that every operator is supported, and test output quality after quantisation.
3. Set up tokenization and the generation loop: sampling, stop conditions, and streaming tokens into the UI without stalling the game.
4. Write the character. Persona, backstory, knowledge and guardrails go into a prompt, and you manage memory as the conversation outgrows the model's context.
5. Add speech-to-text and text-to-speech as two more models to find, convert and run alongside the language model, within the same memory budget.
6. Connect it to gameplay. Parse replies into actions, describe the scene to the model, and drive animation from the character's mood.
7. Repeat for every platform. A model that runs on a desktop GPU may not fit on a phone or a standalone headset.

Each step is solvable. Together they are a machine learning project that runs alongside your game.

On-device or hosted: the trade-offs

On-device characters (Sentis and similar runtimes)

• No per-conversation server cost, and conversations stay on the device.
• Work offline once the models ship with the game.
• Model size, quality and speed are capped by the player's hardware.
• Updating a character means shipping a new build.

Hosted characters (Neocortex)

• Larger models and streaming voice without a hardware budget on the player's side.
• Designers update personality and knowledge in the editor without a new build.
• Need a network connection and a plan that matches your usage.
• Local prototyping is still possible: you can run text chat with local models through Ollama in Unity.

Sentis vs Neocortex at a glance

Why Unity teams pick Neocortex for characters

Skip the ML work. No models to find, convert or tune. Design the character, install the Unity SDK for Unity 6 and open one of the six sample scenes.

Designers own the character. Personality, emotion, knowledge, intents and actions live in a visual graph anyone on the team can edit and test in the browser, without spending credits.

Voice included. Players speak, characters answer in a cloned or designed voice, including in Unity WebGL builds with microphone support.

Replies that drive the game. Each line carries an emotion tag, and actions arrive as data such as { "name": "OPEN_DOOR", "targetId": "door_02" } for your scripts to run.

Use Sentis and Neocortex together

Sentis and Neocortex are not either-or. Use Sentis for on-device perception, such as gesture or pose recognition from the camera, and send what it detects to a Neocortex character as a game event. A player waves, Sentis classifies the gesture, your script reports it, and the character greets them back by voice. Each tool does the part it was built for.

Which to choose

If your goal is a custom model running on the player's hardware, start with Sentis. If your goal is an NPC players can talk to, with a personality your designers control, a voice of its own and actions in your game, start with Neocortex: build the character on the free plan, run it in a sample scene, and add Sentis later for anything you want to detect on-device.

Unity details were last checked in October 2026.

Frequently Asked Questions