Axiomic Labs is an independent, Australia based, AI research lab focused on building efficient, high quality, open language models and high-signal datasets for reasoning, code, and evaluation. Our work explores how small models can achieve stronger capabilities through better training data, compact architectures, custom tokenization, and targeted benchmarks.

Latest releases

Models

• GPT-S-5M - A 5.16M-parameter language model featuring our latest T-X3 architecture and achieving best-in-class performance metrics.
• GPT-X2-125M - Our flagship language model featuring our second-generation T-X2 architecture and achieving near state-of-the-art performance at the 125M scale.

Benchmarks

• ArithMark-2.0 - An updated version of our Arithmetic Benchmark, featuring improvements across the board

Datasets

• NPset-2-Python-Edu - Our latest Natural language code dataset, converted from stack-edu (python split), parsed using the all new NPset-2 spec

On this Hugging Face organization, we share:

• Small language models trained from scratch
• Research datasets for code, math, and reasoning
• Evaluation benchmarks for compact model intelligence
• Open artifacts from our training and experimentation pipeline

We are especially interested in data-efficient pretraining, small-scale reasoning, synthetic evaluation, and pushing architecture forward.

Visit our website: https://www.axiomiclabs.com