Feargal/nvyra-x-reasoning

NRM: Nvyra Recursive Reasoning Model

Developed by Nvyra X — Fact-Checking and Disinformation Detection Service

Model Description

NRM (Nvyra Recursive Reasoning Model) is a state-of-the-art reasoning architecture that combines:

• Mixture of Recursions (MoR) - Weight-tied transformer blocks applied recursively
• Multi-Head Latent Attention (MLA) - 10× KV cache reduction (DeepSeek-V3)
• ConvSwiGLU - Enhanced nonlinearity from URM paper
• Aux-Loss-Free MoE - Bias-based expert load balancing
• PonderNet - Adaptive computation time
• Multi-Token Prediction - 4-ahead planning

Training

• Budget: $115 ($25 Nebius + $90 Modal)
• Hardware: H200 NVLink GPUs
• Framework: PyTorch 2.9.1, Flash Attention 3, CUDA 12.8
• Dataset: 300K+ reasoning examples (Sudoku, ARC, Logic, Object Tracking)

Usage

# This model uses a custom architecture - see repository for full code
from safetensors.torch import load_file
weights = load_file("model.safetensors")

Citation

If you use this model, please cite:

@misc{nrm2025,
  title={NRM: Nvyra Recursive Reasoning Model},
  author={Nvyra X Research Team},
  year={2025},
  url={https://huggingface.co/Feargal/nvyra-x-reasoning}
}

References

• Universal Reasoning Model (URM)
• DeepSeek-V3
• PonderNet