OLMo — Open Language Model

OLMo — Open Language Model

OLMo is the Allen Institute for AI's fully transparent large language model. Unlike most LLMs, OLMo ships the complete stack under Apache-2.0: training code, model weights, pre-training data (Dolma), evaluation harnesses, training logs, and optimizer states. This lets researchers reproduce, scrutinize, and build on the model from first principles.

Key features

• Full reproducibility: all training checkpoints and logs are public
• Integrated with Hugging Face transformers and vllm for inference
• Supports fine-tuning via the companion OLMo-Recipes repo
• Evaluation via lm-eval-harness (AllenAI maintains the harness)
• Multiple scales (1B, 7B, 13B+) with continued updates

Quick start

pip install ai2-olmo

from hf_olmo import OLMoForCausalLM, OLMoTokenizerFast

model = OLMoForCausalLM.from_pretrained("allenai/OLMo-7B")
tokenizer = OLMoTokenizerFast.from_pretrained("allenai/OLMo-7B")

inputs = tokenizer("AI research is", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(out[0]))

npx ai-supply add olmo-open-language-model

Curated mirror of the open-source OLMo (Apache-2.0). Get it from the source.