Tryll Engine Alternative: Cloud, On-Premise or On-Device AI NPCs

A studio is three months into a narrative PC game. The design calls for a companion who remembers the player, reacts to what happens in the world and talks back out loud. The team does not want a server bill for every conversation, so on-device AI looks attractive, and Tryll Engine comes up. Before committing, it helps to know what a Tryll Engine alternative should offer, and what running the AI on the player's GPU means for the rest of the game.

This post covers what Tryll Engine is, the trade-offs of inference on the player's graphics card, how per-copy pricing compares with paying for use, and how Neocortex gives studios cloud, on-premise and on-device options with our team behind the deployment.

Tryll Engine in one paragraph

Tryll Engine is an on-device AI platform for games: its models run on the player's GPU rather than on a server, and it detects NVIDIA, AMD and Intel cards. It covers language models, speech-to-text, text-to-speech, vision-language models and agents. The feature list includes knowledge drawn from game documents and lore, persistent NPC memory and relationships, agents that act in the game world, and runtime generation of quests, lore and events. A model manager picks a quantized open model sized to the player's hardware, from small models of a few billion parameters up to around 14B. Unity and Unreal plugins are in early access on request, a native C++ route is listed as planned, and the documentation is a pre-release that expects breaking changes before a 1.0 API. It needs 8GB of VRAM at minimum and recommends 12GB or more. It is free to start, then charges about $0.08 per copy sold on a $15 game, scaling with the game's price. It is backed by Epic MegaGrants and Early Game Ventures, and its site lists no shipped games.

What running AI on the player's GPU asks of your game

On-device inference takes the server out of the conversation. It also moves the cost of the model onto the player's machine, and that cost shows up in four places. The same applies to any library that runs models beside your game, as our LLMUnity comparison and NVIDIA ACE comparison describe.

A hardware floor

An 8GB minimum with a 12GB recommendation decides who gets the feature. Players below the floor either lose the AI characters or get a smaller model. For a game whose story leans on a companion, that is a design decision, not a settings toggle: every scene the character carries needs a fallback.

Sharing the frame budget

The GPU that renders your game is the GPU that runs the model. VRAM is split between textures and model weights, and token generation competes with rendering for compute. Tryll's guidance pairs AAA games at high settings with a 1B to 3B assistant and keeps 7B and up for lighter games such as visual novels. The richer your graphics, the smaller the character's model.

Smaller models, different writing

Models that fit beside a game are smaller than hosted ones. They can hold a persona, but they need tighter prompts, shorter context and more guardrails, and quality drops sooner as conversations grow. Writers end up tuning prompts for each model size the manager might pick.

Testing per device

Every GPU tier, driver version and quantization level is another combination to test. A line that reads well on a 16GB card may come out differently from the smaller model a laptop receives. The QA plan grows with the hardware matrix, and so does the support queue after launch.

None of this rules out on-device AI. It means the choice of where inference runs belongs to each project, and one deployment model will not suit every game.

Per-copy pricing vs paying for use

Tryll charges a share of each copy sold. At about $0.08 on a $15 game, it costs nothing until the game sells, and the bill follows sales rather than playtime. The fee applies to every copy, including players who never speak to a character, and it rises with your price point.

The self-serve Neocortex product is priced by plan:

• Free: 2 characters and 200 monthly credits, for personal use.
• Starter: EUR 9 a month, for personal and academic use.
• Pro: EUR 29 a month, with commercial use, group chat and 5 custom voices.
• Team: EUR 99 a month, with 5 seats, the quiz engine and 20 custom voices.

Testing characters in the browser editor never uses credits. On-premise and on-device deployments are Custom arrangements, scoped in a conversation with our team around your hardware and audience. For a wider view of how vendors charge, see our AI NPC pricing models comparison.

A Tryll Engine alternative with three ways to deploy

Neocortex is a platform for AI characters in games. You design a character in a visual node editor, with nodes for personality, emotion, knowledge, world, objects, intents, actions, language, accent and voice, then connect it to your game through SDKs for Unity, Unreal Engine 5 and Godot (beta), or through the Web API from any language. Where the character runs is the studio's call.

Cloud. The standard self-serve product. The model runs on our servers, so the player's GPU stays with your graphics and every player meets the same character whatever their hardware. It is built for real-time conversation, with streaming speech-to-text and text-to-speech, voice cloning from a short sample you have the rights to, and voice design from a text description.

On-premise. The full platform on your own servers or private cloud: the character editor, the API, your characters and Conversation Flows. Speech-to-text and text-to-speech run on-premise too, not only the language model. It can run fully air-gapped with no internet connection, for secure sites, training simulators, kiosks and events. Deployments are sized with the studio.

On-device. Inference can ship inside the game build and run on the player's machine, for projects where local play is the requirement.

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.

On-premise and on-device are Custom (enterprise) arrangements. Our team works with the studio on deployment, model choice and integration, so the hardware floor, the model size and the fallback plan are settled before launch instead of discovered after it. To talk through a local or on-premise setup, contact us.

The character design carries across. Conversation Flows return flowState checkpoint keywords your game reacts to, and replies carry an emotion tag plus actions such as { "name": "OPEN_DOOR", "targetId": "door_02" } for your scripts to run.

Tryll Engine vs Neocortex at a glance

Next steps

1. Decide where your characters must run. If every player needs them, a VRAM floor is a design risk. If the build must work offline or on a closed network, plan for local inference from the start.
2. Prototype the character now. Build it in the Neocortex editor on the free plan and test it in the browser. The design carries over to whichever deployment you choose.
3. See the wider field. Our roundup of AI NPC tools with local AI support and the comparison hub cover the other routes.
4. Talk to our team about local deployment. If the project needs on-premise, air-gapped or on-device characters, we scope the hardware, the models and the integration with you.

Tryll Engine details checked October 2026.

Frequently Asked Questions