Repository navigation
BehaviorLLM 0.6.0 — local model-driven decisions in Unity
BehaviorLLM lets Unity characters and game systems choose from actions you define, using a local language model. You supply observations and bind actions to game code; the package builds a constrained request, validates the reply, and dispatches the decision.
Included
- StealthGuard and PrisonYard samples, with three and eight decision makers respectively.
- Model Manager for discovering, downloading and activating compatible GGUF models and installing the local backend.
- Automatic default configuration assets, with custom assets supported.
- Prompt Inspector showing requests, schemas and replies.
- Named action parameters, runtime action availability, and argument validation.
- Sample builders under BehaviorLLM > Samples, including native PrisonYard models and animations without a glTFast dependency. Primitive fallback remains available.
Install
Requires Unity 6, a local llama-server backend and a compatible GGUF model. The gameplay samples additionally require AI Navigation and Input System.
Use Unity Package Manager > Add package from git URL:
https://github.com/faelion/BehaviorLLM.git#v0.6.0
Open BehaviorLLM > Model Manager, install and activate a model, then build a sample from BehaviorLLM > Samples.
Migration from 0.5
Breaking API change: action handlers receive ActionArguments instead of string. Use args.First for a single value, or args["name"] / args.GetInt("name") for named parameters. Existing serialized single-argument definitions migrate to a parameter named arg; review custom handlers and event bindings.
Validation and limitations
Release validation on September 22 passed 135 EditMode and 137 PlayMode tests on Unity 6000.3.1f1 in a fresh consumer installed from GitHub. Both shipped scenes match their builders, including native art, animation and all six prison section volumes. Two live smoke tests with local Qwen3.5-2B produced model-selected actions for all 11 decision makers: 17 decisions, zero fallbacks and zero backend errors. This is a short integration check, not a new performance benchmark. Primitive fallback was separately validated on September 21.
The historical model matrix yielded 192/192 structurally valid replies. That does not imply correct decisions: schema-enabled expected-action rates varied by configuration. Current-code reruns are included under Experiments~/runs/; latency and accuracy depend on the model, workload and hardware. No new standalone-player certification is claimed.
Jev and openjev are not integrated in this release. Decision-model backends are a possible future experiment.
MIT-licensed package; bundled sample art retains its own licenses. See the README, changelog and sample license files for details.
Feedback
Which would help your project most: more game-system examples, better decision debugging, or an experimental local decision-model backend? Share your Unity version, model and hardware when reporting a problem. Jev support is planned future work, not part of 0.6.0.