aimar00/4x-mamba

4X Mamba

Legacy public 4X strategy-game Mamba checkpoint.

This is the first model: a pretrained Mamba2-style sequence model trained with a next-token prediction objective over synthetic 4X strategy-game traces. It is kept as a reference checkpoint, not as the current latent world model.

The repository has been cleaned so the public artifact is weights-only:

• model.safetensors - model weights only
• config.json - minimal model/checkpoint metadata

No optimizer state, scheduler state, training corpus, simulator source, infrastructure scripts, benchmark harness, or unrelated project files are included in the cleaned artifact.

Clean public code for the current world model lives here:

https://github.com/ProjectAI00/4x-mamba

Model

• Architecture: legacy Mamba2-style sequence model
• Objective: next-token prediction
• Parameters: 27,648,512
• Original training metadata:
  • Global step: 47,500
  • Epoch: 3
  • Validation loss: 0.7532
  • Validation accuracy: 0.8130

The current latent transition world model is published separately at aimar00/4x-mamba-world-model.