AI2 releases OLMo 2 32B - fully open model with training data, code, and weights under Apache 2.0
The Allen Institute for AI released OLMo 2 32B, a 32B dense model with everything open: training data, training code, model weights, and evaluation tools, all under Apache 2.0. It is the most transparent large model release to date.
The Allen Institute for AI (AI2) released OLMo 2 32B on April 3, continuing their commitment to fully open AI research. Unlike most model releases that share only the final weights, OLMo 2 opens everything: the weights, the training code, the training data, intermediate checkpoints, and evaluation tools.
The release includes the complete Dolma training dataset, which is documented and reproducible. Every design decision is recorded, from architecture choices to hyperparameter tuning. All of it is under the Apache 2.0 license with no restrictions, no revenue thresholds, and no usage limitations whatsoever.
At 32B parameters with a dense architecture, OLMo 2 is not the most capable model released in April. It does not match the benchmark scores of Qwen3-72B or Llama 4 Scout on most tasks. But raw capability is not the point. The point is reproducibility and transparency - knowing exactly what went into the model and being able to verify every claim about it.
If you need to explain to a regulator, client, or auditor exactly how your AI model was trained, what data it saw, and what biases it might have, OLMo 2 is the only large model where you can actually answer those questions completely. Every other "open" model release leaves gaps in the documentation.
The model runs on practical hardware. At 4-bit quantization, it needs approximately 20GB of VRAM, fitting on a single RTX 3090. Inference speed is reasonable for interactive use, comparable to other dense 32B models. Ollama and vLLM both support it out of the box.
For researchers and organizations building on top of open models, OLMo 2 serves as a reference implementation. You can study how a 32B model is trained from scratch, experiment with different training approaches using the same data, and benchmark your own modifications against the original results.
AI2's approach challenges the industry norm where "open-weight" means "here are the weights, good luck figuring out the rest." Full transparency takes more effort to produce, but it creates models that the community can truly build on, verify, and improve. With the EU AI Act's transparency requirements taking effect, this level of openness may become a competitive advantage for model providers.
