Furuta/Termetry-hrm-text-v0.1

Termetry HRM-Text v0.1

Unofficial experimental release.

Termetry HRM-Text v0.1 is an experimental Apple Silicon model based on HRM-Text and inspired by the Bonsai / 1.58-bit ternary model direction.

The goal of this release is simple: make a small HRM-Text variant that can be tested locally with MLX.

References

• Termetry HRM-Text v0.1: https://huggingface.co/Furuta/Termetry-hrm-text-v0.1
• Base model, HRM-Text-1B: https://huggingface.co/sapientinc/HRM-Text-1B
• HRM-Text paper, "HRM-Text: Efficient Pretraining Beyond Scaling": https://arxiv.org/abs/2605.20613
• HRM-Text code: https://github.com/sapientinc/HRM-Text
• PrismML Bonsai announcement: https://prismml.com/news/bonsai-8b
• Bonsai whitepaper: https://github.com/PrismML-Eng/Bonsai-demo/blob/main/1-bit-bonsai-8b-whitepaper.pdf
• Ternary Bonsai 8B MLX 2-bit: https://huggingface.co/prism-ml/Ternary-Bonsai-8B-mlx-2bit
• 1.58-bit / BitNet b1.58 paper, "The Era of 1-bit LLMs": https://arxiv.org/abs/2402.17764

Notes

• Base: sapientinc/HRM-Text-1B
• Status: unofficial experimental variant
• Runtime target: MLX on Apple Silicon
• Weight format: safetensors
• Quantization: 2-bit affine, group size 128
• Ternary direction: {-1, 0, +1} style low-bit weights

This is a v0.1 experimental release. Evaluation results and cleaner usage examples will be added in later versions.

This release does not guarantee the same quality, accuracy, benchmark scores, or behavior as the original HRM-Text-1B model or the referenced Bonsai models. It should be treated as an experimental low-bit variant for local testing.

License

The base model is published with Apache-2.0 metadata, and this release keeps the same license metadata.