trina

# Trina Trina is a highly-performant, self-contained LLM AI plugin for Paper/Spigot Minecraft servers. It uses Panama Vector math operations in Java 21 to run lightweight local LLMs directly in your Minecraft server's resources without needing external python scripts or GPUs. ## Requirements - Java 21 or newer. - Minecraft Paper/Spigot server (tested only on 26.1.2!) - The JVM must be launched with incubator modules enabled: `java --add-modules=jdk.incubator.vector --enable-preview -jar server.jar` - ~2GB to ~6GB of RAM depending on the selected model. ## Building To compile the plugin from source, simply run the included Gradle wrapper: ```bash ./gradlew shadowJar ``` ## Configuration When you first run the plugin, it will generate a `config.yml` in the `plugins/Trina/` directory: ```yaml # Select the model to load (1-9): # 1 : Qwen 0.5B ~500MB # 2 : Llama 3.2 1B ~1.2GB (default, best balance) # 3 : Yi-Coder 1.5B ~1.5GB # 4 : Gemma 2 2B ~2GB # 5 : Granite 3.0 2B ~2GB # 6 : Llama 3.2 3B ~2.5GB # 7 : Mistral 7B ~5GB # 8 : Llama 3.1 8B ~5.5GB # 9 : Yi-Coder 9B ~6GB (best for code) model: 2 # Number of CPU threads to use for generation (0 = auto-detect) # Changing this config requires a server restart threads: 0 # Allow AI to run server commands via the Model Context Protocol allow-mcp: true system_prompt: "You are a helpful AI assistant." temperature: 0.7 max_tokens: 256 # Web UI and OpenAI-compatible API hosting host: 0.0.0.0 port: 8080 # Optional Bearer token for the /v1/chat/completions endpoint (leave empty for none) api_key: "" ``` ## Legal & Credits [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0) This project operates around the incredible [Jlama](https://github.com/tjake/jlama) inference engine.