OpenLanguageModel

OpenLanguageModel (OLM) is an MIT-licensed Python library for building and training transformer language models in PyTorch. It is written for people who want to read and modify the model code rather than call a black box: architectures are assembled from plain, named PyTorch modules - attention, normalisation, feed-forward, routing - so a model definition stays close to the equations it implements.

The package ships reference implementations of 27 architectures with 47 named size presets, covering GPT-2, Llama 2, 3 and 4, Qwen 2.5, 3 and 3.5, Gemma 3, Phi-3 and Phi-4, Olmo, DeepSeek-V3, Kimi K2, MiniMax M2 and Step 3.5. Training runs on CPU, on a single GPU, or on one node with several GPUs through DDP or FSDP. It also includes BPE tokenizer training, dataset tooling, a conceptual learning path in the documentation, and four Colab notebooks.

It installs from PyPI and supports Python 3.10 to 3.12. The package is still classified Alpha, and it is a research and teaching library rather than a serving or deployment stack.