---
license: other
library_name: diffsynth
tags: [minecraft, world-model, video-generation, inverse-dynamics, agent]
---

# EvoWAM — wild-mine checkpoints

Custom checkpoints for the **WAN + IDM closed-loop Minecraft agent** (acts by imagining the next ~1s
of gameplay with a video model and decoding it into controls, closed-loop on the real game).
See the GitHub repo `release_wan_idm_wildmine/` for the runnable code (`run_closed_loop.py`).

| file | size | what |
|---|---|---|
| `wan_fullft/step-52834.safetensors` | 9.4 GB | full fine-tune of the **Wan2.2-TI2V-5B** DiT on Minecraft gameplay + VPT-style action captions |
| `vpt/4x_idm.model`, `vpt/4x_idm.weights` | ~1.8 GB | **OpenAI VPT** 4x Inverse Dynamics Model (re-hosted for convenience) |

## Base models (NOT in this repo — download from source)
- WAN-2.2-TI2V-5B: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
- Qwen3-VL-8B-Instruct: https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct

## Usage
```
huggingface-cli download TMarcus/EvoWAM-wildmine-checkpoints --local-dir ./ckpts
# then point the runner at them:
#   --wan-fullft ckpts/wan_fullft/step-52834.safetensors
#   --idm-model ckpts/vpt/4x_idm.model  --idm-weights ckpts/vpt/4x_idm.weights
```

Licenses: the fine-tune derives from Wan2.2-TI2V-5B (see Wan-AI's license); the IDM is OpenAI VPT.