Meta releases the OPT family with restricted access to OPT-175B
Meta releases the 125M-to-175B OPT family; smaller checkpoints and code are available, while OPT-175B is limited to approved noncommercial research access.
What happened
Event details
On May 3, 2022, Meta released the OPT family and announced OPT-175B. The paper said the 175B model was broadly comparable to GPT-3 and disclosed detailed training logs, infrastructure failures, and training code. Compared with API-only models, this gave outside researchers substantially more material to inspect. The release was not fully open. Smaller checkpoints and code were available directly, but OPT-175B used a noncommercial license and required manually reviewed applications. In a follow-up two months later, Meta said it had received more than 4,500 requests and granted access to 668 entities across 49 countries. Those figures show both strong research demand and the practical limits of the initial access program.
Assessment
Why it matters
OPT brought weights, training code, and detailed logs for a 175B-scale model into external research, improving the inspectability of large-model training. Its main influence was on research transparency and staged release practice. The noncommercial license and manual approval process constrained reproduction and commercial use, so it should not be described as fully open source.
Availability
Access notes
Smaller checkpoints and code were available; OPT-175B was limited to approved institutions for noncommercial research.