Mica v0.1 4B: Crafting an Iron Pickaxe in Real Minecraft Without Generating a Single Token

A 4-billion-parameter model crafted an iron pickaxe from an empty inventory in a real Minecraft 1.20.4 server — without ever generating a single text token. The project, called Mica v0.1 4B, sparked discussion on Reddit for demonstrating an agent control approach fundamentally different from the mainstream paradigm of "having LLMs output action text."

Core Mechanism: Scoring Decisions Instead of Generating Them

Most game agents work by having a language model "speak" — outputting the next command as text, which a parser then converts into game actions. Mica takes a different route.

Every step follows the same loop: the robot's real-time game state (inventory items, nearby blocks, surrounding entities, result of the previous action) is transcribed into text; Mica then scores a set of candidate commands and selects the next action; finally, Mindcraft's skill library (a Mineflayer-based robot framework) executes the chosen command in-game.

The key is in the scoring step. Mica never generates text — instead, it directly reads the probability values of answer-label tokens to rank candidate commands and pick the best one. In other words, the model outputs zero tokens — it doesn't write anything; it just computes probabilities. This "classification-style" rather than "generation-style" decision-making compresses an open-ended generation problem into a constrained selection problem.

The technical background of reading token probabilities directly is worth unpacking. During inference, a language model outputs a logit value for every token in its vocabulary; after softmax normalization, these become a probability distribution. Mica operates at exactly this layer: given a prompt describing the current state, the model is asked to "choose the next command," but never actually decodes into full text — instead, it directly extracts the probability values of tokens corresponding to preset candidate commands (the "answer labels") and ranks by those. This approach is sometimes called "constrained decoding" or "classification-style inference," and has seen use in intent classification and multiple-choice evaluation, though using it for real-time action selection in game agents remains a niche approach. Its core advantages are: the candidate action set is closed, eliminating the possibility of malformed output or hallucinated commands; and since only a single forward pass is needed to read specific token logits rather than autoregressively generating step by step, inference overhead is dramatically reduced.

The Full 23-Step Path to an Iron Pickaxe

The project author provided a complete run log: starting from an empty inventory, Mica completed the entire workflow from nothing to an iron pickaxe in 23 decisions.

This chain covers the full standard early-game progression in Minecraft: chopping logs, crafting planks, making a crafting table, building a wooden pickaxe, mining stone, upgrading to a stone pickaxe, building a furnace, mining iron ore, smelting, and finally crafting the iron pickaxe. For anyone familiar with the game, this is a sequence that requires correct ordering, where each step depends on the output of the previous one — a single wrong step can stall the entire pipeline.

Stably completing this in 23 steps demonstrates practical-level reliability in state understanding and task planning, rather than a lucky one-off success.

Performance: Real-Time Decisions on a Single GPU

The runtime specs are equally noteworthy. Mica runs on llama.cpp with Q5_K_M quantization on a single RTX 3090.

Under this configuration, each decision takes approximately 90 to 150 milliseconds. This latency means the agent can respond to game state changes in near real-time, without waiting seconds per inference step. For a real-time game environment requiring continuous decisions, this response speed is the dividing line between usable and not.

It's worth emphasizing that with 4 billion parameters plus quantization, the entire system runs on a consumer-grade GPU rather than relying on large-scale clusters — lowering the barrier to reproduction and further development.

Q5_K_M is a mixed quantization scheme in the GGUF format, introduced by the llama.cpp project. It applies different precision levels to different layers (some at 5-bit, others at lower precision), compressing VRAM usage while preserving model quality as much as possible — generally considered a good quality-to-size tradeoff. For a 4B parameter model, full FP16 precision requires roughly 8GB of VRAM, while Q5_K_M compresses this to around 3GB, easily fitting on a consumer GPU. llama.cpp itself is a pure C/C++ inference framework heavily optimized for CPU and GPU hybrid inference, and is currently one of the most widely used toolchains for local deployment of open-source models. Together, they allow Mica's entire system to run on an ordinary gaming PC without needing large-scale CUDA parallel clusters.

Visualization: Every Decision Is Transparent and Auditable

The demo video the author provides puts considerable effort into presentation. A right-side panel displays the full decision process in real time for every step: what the candidate commands are, the probability distribution Mica assigns, the final selection, and the execution result. Each step is simultaneously logged in a history feed.

Long actions (such as walking, mining, and smelting) are sped up in the video with playback speed labels, and sequences of consecutive retries are compressed in editing. The HUD interface, crafting table, and furnace views are rendered based on inventory state logged by the robot.

This transparent presentation lets observers intuitively understand why the model made each choice, rather than just seeing a black-box output — which is highly valuable for evaluating agent decision quality.

Significance and Open Resources

Mica's approach offers a valuable reference: for tasks with relatively limited action spaces where candidates can be enumerated, replacing text generation with probability scoring can reduce inference overhead while avoiding the formatting errors and hallucination problems of generative outputs. This has practical implications for building lightweight, reliable game agents or tool-calling systems.

As a v0.1 release, it remains in early stages, and a single 23-step demo is insufficient to demonstrate robustness on more complex, longer-horizon tasks. But the author has open-sourced both the model weights and full code:

• Model weights: sky7350/Mica-v0.1-4B on Hugging Face
• Code and server: akivet/Mica-v0.1-4B on GitHub

Being open-source means the community can directly verify how well this "zero-token generation" decision paradigm transfers to other tasks — which may ultimately prove more valuable than a single successful Minecraft demonstration.

Mineflayer is a Node.js-based Minecraft bot development framework that provides low-level game protocol wrappers, supporting API calls for movement, mining, crafting, and interaction. Mindcraft builds on top of Mineflayer to provide a higher-level "skill library," wrapping common game tasks (like "chop tree," "craft planks," "go to coordinates") into directly callable functions. The candidate commands Mica selects from are precisely these predefined skills, not low-level atomic operations. This layered architecture means Mica's decision layer only needs to choose among a finite set of high-level skills, dramatically reducing action space complexity — and is one of the prerequisites for completing a long workflow chain with zero token generation. Understanding this helps clarify the method's applicable boundaries: the more closed and structured the action space, the more pronounced the advantage of scoring over generation; conversely, if a task requires open-ended planning or generating new skills, the limitations of this paradigm will become apparent.