Before ChatGPT, There Was GPT-1
Long before "GPT" became a household abbreviation, OpenAI's original GPT -- released in June 2018 -- was a modest research paper and a comparatively small checkpoint. It introduced the now-familiar recipe of unsupervised pretraining on a large text corpus followed by fine-tuning for specific tasks, built on the recently-published Transformer architecture.
Why it mattered
GPT-1 wasn't the most powerful language model of its era in absolute terms, but it set the template every subsequent GPT release -- and much of the field -- would follow: a decoder-only Transformer scaled up and pretrained generatively. Every major LLM released since, open or closed, sits downstream of the approach this release popularized.
It's also a useful historical marker for this registry's own premise: OpenAI's later releases became progressively more closed (GPT-2's staged rollout, then GPT-3 onward as API-only), making this early, openly released checkpoint one of the last direct downloads from OpenAI available to hash at all until 2025's GPT-OSS-20b.
On the record
- Issuer: OpenAI
- License: MIT
- Claimed cutoff: 2018-06-11
- Status: on-chain, attested
See the full hash breakdown and on-chain attestation on GPT-1's Explorer page.