I think it’s a big deal to see a fully open LLM now achieving this level of quality. While the partially open releases we’ve seen from the big labs are are quite valuable, models like OLMo-2 are the only way that researchers can truly study this technology to answer questions about how the models’ capabilities are shaped by their training data and training process.

The closed and partly-closed models rely on a lot of secret sauce, so it’s also just really impressive to see their results being replicated in the open.