Mindcraft-CE/Andy-4.2
Mindcraft-CE/Andy-4.2 is a 9 billion parameter model developed by Mindcraft CE, built on the Qwen3.5 architecture with Gated Deltanet attention. It is specifically optimized for playing Minecraft, demonstrating advanced in-game comprehension and the ability to achieve complex objectives like acquiring full diamond armor autonomously. This model is designed for local deployment, capable of running on a single RTX 3090 with 256k context at 8-bit quantization, and supports up to 1 million tokens of context.
Loading preview...
Mindcraft-CE/Andy-4.2 Overview
Andy-4.2 is a 9 billion parameter model from Mindcraft CE, designed as a highly capable local AI for playing Minecraft. It leverages a new Qwen3.5 architecture with Gated Deltanet attention, enhancing both intelligence and efficiency compared to previous versions. This allows it to run effectively on a single RTX 3090, supporting up to 1 million tokens of context and stable operation for 120 messages.
Key Capabilities
- Autonomous Minecraft Gameplay: Andy-4.2 is the first local model demonstrated to achieve a full set of diamond armor in Minecraft with zero human interaction.
- Enhanced Vision: Features a stronger multimodal base for deeper comprehension of the game state.
- Efficiency: The Gated Deltanet architecture enables efficient local execution, even with 256k tokens of context at 8-bit quantization.
- DeepSeek-R1 CoT Style: Utilizes a DeepSeek-R1 style Chain-of-Thought for improved reasoning.
Limitations and Use Cases
While excelling in Minecraft navigation and objective completion, Andy-4.2 has a known limitation in building tasks; it is not recommended for code generation or complex building actions. The model was trained using 4-bit QLoRA with 8-bit QaT on 2,748 examples over 5 hours. It is recommended to run Andy-4.2 using LM Studio for optimal performance, with default settings proving effective for its core capabilities.