The Mindcraft CE team introduces Andy-4.2 Air, noted as the best local AI you can use to play Minecraft with. Thinking faster than Andy-4.1, being able to carry out more actions than Andy-4, and rivaling models 10x it's size.
Key Innovations
Andy-4.2 Air uses largely the same formula as Andy-4.1, but introduces a new architecture from the Qwen3.5 series which makes the model not only smarter, but more efficient. Using Gated Deltanet attention allows Andy-4.2 to run on a single RTX 3090, with 256k tokens of context, at a staggering 8-bit quantization.
Andy-4.2 Air is also the first local model capable of getting a full set of diamond armour, with zero human interaction
Andy 4.2 Air with Diamond Armour
Like Andy-4.1, Andy-4.2 has vision capabilities, and has a stronger multimodal base that allows for even deeper comprehension of the game state.
How to Run
Andy-4.2 Air is still recommended to be ran using LM Studio, we have tried using Ollama, and there were a plethora of issues, including looping, mismatched chat templates, etc;
Below are the recommended sampling parameters for Andy-4.2 Air, but the default settings in LM Studio work great, and the model is still able to get full diamond armour by itself.
Parameter
Value
Temperature
0.6
Top K
20
Repeat Penalty
1.0
Top P
0.95
Min P
0
Model Specifications
Size: 4B Parameters
Architecture: Qwen3.5
Context Length: Up to 1 million tokens
Message Count: 120 messages stable
CoT Style: DeepSeek-R1 style
Training Specifications
Hardware: 1x RTX 3090
Training Time: 3 Hours
Dataset Size: 2,748 examples
Learning Rate: 2e-5
LR Scheduler:cosine
Epoch Count: 1 Epoch
Training Quantization: 8-bit LoRA with 8-bit QaT
Testing
The testing for Andy-4.2 Air was done at 16-bit.
Andy-4.2 Air had the following stats during runtime for testing:
Mindcraft-CE version 1.2.7
16-bit Quantization
16-bit KV Cache quantization
Base LM Studio sampling parameters
32,000 Context Length
65 Messages
Limitations
Even though Andy-4.2 Air is capable of incredible feats, there is one domain where it does not perform well: Building. During internal testing any time Andy-4.2 would use !newAction, it would produce thousands and thousands of tokens, but never do anything. It is not advised to use Andy-4.2 as your code model.
Apart from that. Andy-4.2 Air has shown to be our most hard-working model yet, and navigates potential errors very well.
What's Next?
Based on the lessons from Andy-4.2 models, the Mindcraft team is prepared to collect better training data, explore new architectures to make the cost of running Andy models cheaper, as well as packing more brains into these tiny minds.
Licenses and Notices
Like all other Andy models, Andy-4.2 Air is based on the Andy license of terms. Being generally permissive, it contains qualifiers as to what makes an "Andy" class model.
