Mica v0.1 4B

Mica v0.1 4B is a 4-billion-parameter model capable of progressing from an empty inventory to an iron pickaxe on a real Minecraft 1.20.4 server without generating any text tokens.

Instead of generating text, Mica directly reads the probability values of answer-label tokens, uses them to rank candidate commands, and selects the optimal one, with the model outputting zero tokens.

Mica's decision loop consists of: transcribing the game state into text, scoring a set of candidate commands to select an action, and then having the Mindcraft skill library execute the selected command in the game.

This classificatory rather than generative decision-making approach compresses the open-ended generation problem into a constrained selection problem, with a closed candidate action set that eliminates output formatting errors and hallucinated commands.

Mica completed the entire workflow from empty inventory to iron pickaxe in 23 decisions, covering chopping logs, crafting planks, making a crafting table, wooden pickaxe, mining stone, stone pickaxe, building a furnace, mining iron ore, smelting, and crafting an iron pickaxe.

Mica runs on llama.cpp with Q5_K_M quantization on a single RTX 3090.

Under this configuration, each of Mica's decisions takes approximately 90 to 150 milliseconds, enabling near-real-time responses to changes in game state.

Mica's model weights are published on Hugging Face at sky7350/Mica-v0.1-4B, and the code and server are published on GitHub at akivet/Mica-v0.1-4B.

Mindcraft builds a higher-level skill library on top of Mineflayer, encapsulating common game tasks into directly callable functions; Mica's candidate commands are these predefined skills.

As an early v0.1 release, a single 23-step demonstration is insufficient to prove the model's robustness on more complex, longer-horizon tasks.