---
license: other
license_name: lfm1.0
license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE
language:
  - en
base_model: LiquidAI/LFM2.5-350M
datasets:
  - DedeProGames/Andy-4.1-NanoAndy
library_name: transformers
pipeline_tag: text-generation
tags:
  - minecraft
  - mindcraft
  - mindcraft-ce
  - agent
  - lfm2
  - edge
  - conversational
---

![NanoAndy-350M](logo.png)

# NanoAndy-350M

**NanoAndy-350M** is a compact, full-fine-tuned Minecraft agent model built on [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M). It is inspired by [Andy-4.1](https://huggingface.co/datasets/Mindcraft-CE/Andy-4.1) and purpose-built to run as the "brain" of a bot in [Mindcraft-CE](https://github.com/mindcraft-ce/mindcraft-ce), the community fork of the open-source Mindcraft platform that lets LLMs control Minecraft characters via Mineflayer.

At only 350M parameters, NanoAndy-350M is meant for setups where a full-size Andy-4 model isn't practical — low-VRAM GPUs, CPU-only machines, or running many bots at once.

## What makes it "Nano"

NanoAndy-350M is trained on a stripped-down version of Andy-4.1's conversational data ([DedeProGames/Andy-4.1-NanoAndy](https://huggingface.co/datasets/DedeProGames/Andy-4.1-NanoAndy)):

- **No chain-of-thought.** All `<think>...</think>` reasoning traces were removed from the assistant turns. A 350M model has little spare capacity for long internal monologue, so training goes straight to the final in-game response.
- **No function-calling turns.** Conversations that used external `tool`-role calls were dropped entirely, keeping the model focused on Mindcraft's native chat/command format instead of a JSON tool-calling schema it would rarely use well at this size.

The result is a lean, fast, direct-response model rather than a smaller reasoning model.

## Model Details

| | |
|---|---|
| **Base model** | [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M) |
| **Architecture** | LFM2 (hybrid conv + attention) |
| **Parameters** | ~354M |
| **Fine-tuning method** | Full fine-tune (no LoRA/adapters) |
| **Context length** | 11,264 tokens |
| **Language** | English |
| **License** | [LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE) (inherited from base model) |

## Training

| | |
|---|---|
| **Dataset** | [DedeProGames/Andy-4.1-NanoAndy](https://huggingface.co/datasets/DedeProGames/Andy-4.1-NanoAndy) (1,695 conversations) |
| **Framework** | [Unsloth](https://github.com/unslothai/unsloth) |
| **Hardware** | 1x NVIDIA T4 (Google Colab) |
| **Epochs** | 2 |
| **Effective batch size** | 8 (1 x 8 grad. accumulation) |
| **Learning rate** | 5e-5, cosine schedule, 15 warmup steps |
| **Optimizer** | adamw_8bit |
| **Final train loss** | ~0.39 |

## Usage

NanoAndy-350M is meant to be dropped into a Mindcraft-CE [bot profile](https://github.com/mindcraft-ce/mindcraft-ce) (e.g. `andy.json`) as the chat/coding model, served locally through something like LM Studio, llama.cpp, or vLLM.

It also works with standard `transformers`:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "DedeProGames/NanoAndy-350M"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "You are a minecraft bot named Andy. A player asks you to gather 4 oak logs."
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
)["input_ids"].to(model.device)

model.generate(
    input_ids,
    do_sample=True,
    temperature=0.3,
    repetition_penalty=1.05,
    max_new_tokens=256,
    streamer=streamer,
)
```

## Limitations

- **Not a reasoning model.** With chain-of-thought training data removed, NanoAndy-350M won't show its work — it goes straight to an action/response.
- **No native tool-calling.** It was not trained on function-call syntax; it expects Mindcraft's native command/chat format.
- **Small model.** At 350M parameters, it will struggle with long-horizon planning, complex builds, and multi-step reasoning compared to larger Andy-4 variants.
- **English only.**
- Narrowly tuned for the Mindcraft agent format — not intended as a general-purpose assistant.

## Acknowledgements

- [Liquid AI](https://liquid.ai) for the LFM2.5 base model.
- [Mindcraft-CE](https://github.com/mindcraft-ce/mindcraft-ce) and the Andy-4.1 dataset authors for the source conversational data and the platform this model targets.
- [Unsloth](https://github.com/unslothai/unsloth) for the training tooling.