Echo-1.58 v1.0

First public release of Echo-1.58: native ternary (1.58-bit) quantization-aware training for masked-diffusion language models.

Includes the training/evaluation code, the per-run result tables behind every number in the paper, and the paper PDF. Model checkpoints (four 341M models and the 27M research ladder) are on the Hugging Face collection.

Paper: Recovering the Ternary Penalty in Masked Diffusion Language Models: What Adds Capacity, What Only Reorganizes, and What Repetition Costs.