Xaedro

Independent AI research / Argentina

Reasoning systems
and machine learning
research.

XAEDRO is an independent AI research and engineering organization based in Argentina.

Research GitHub

Featured research / October 2026

Hakuli

An independent clean-room implementation and reproducibility study of APOLLO for memory-efficient LLM training on a single consumer GPU.

Hakuli follows the published APOLLO method and makes no claim of algorithmic novelty. In a matched black-box comparison, Hakuli and the official implementation track closely over 1,000 optimizer steps.

Technical report / Evaristo Cordoba Ancajima / XAEDRO
PyTorch-only / Apache License 2.0

Exact logged series
LLaMA-7B / batch 16 / seq 256 / RTX 3090

Training and validation loss over 1,000 optimizer steps Exact training and validation loss series for Hakuli and the official APOLLO implementation over 1,000 optimizer steps. LLaMA-7B / BATCH 16 / SEQ 256 / RTX 3090 FIGURE 01 2.8 2.4 2.0 1.8 0 200 400 600 800 1,000 OPTIMIZER STEP CROSS-ENTROPY LOSS SOLID / TRAIN     DASHED + DOT / VALIDATION FINAL VALIDATION / HAKULI 2.2055    APOLLO 2.1660
Plotted from the reported training and validation logs

Company

Independent research
and engineering.

XAEDRO is an independent AI research and engineering organization based in Argentina. We work on reasoning systems, evaluation, efficient training, and machine learning systems.

research@xaedro.com contact@xaedro.com