AI2 Releases OLMo 2 32B: Pushing the Boundaries of Open-Source Language Models

Valeriia Kuka

March 25, 2025

2 minutes

🟢easy Reading Level

The Allen Institute for Artificial Intelligence (AI2) has announced OLMo 2 32B. It is notable for being the first fully open model—with all data, code, weights, and methodological details freely available—to outperform commercial models like OpenAI's GPT-3.5-Turbo and GPT-4o mini on a comprehensive suite of academic benchmarks.

Technical Specifications and Performance

OLMo 2 32B represents the largest iteration in the OLMo 2 family, which also includes 7B and 13B parameter versions. The 32B model was trained on up to 6 trillion tokens through a multi-stage process that included pretraining, mid-training, and post-training optimization.

According to AI2's benchmarking, OLMo 2 32B not only matches or outperforms GPT-3.5 Turbo and GPT-4o mini, but also competes with other leading open-weight models including Qwen 2.5 32B and Mistral 24B. The model approaches the performance of much larger models like Qwen 2.5 72B and Llama 3.1/3.3 70B, despite having fewer parameters.

A key achievement is the model's computational efficiency. AI2 reports that training OLMo 2 32B required only about one-third of the computational resources needed for Qwen 2.5 32B while achieving similar performance levels.

Implications for the AI Research Community

The complete open-source nature of OLMo 2 32B provides the research community with unprecedented access to study language model development. This transparency enables investigations into pretraining dynamics, the impact of data on model behavior, and how different training stages interact.