AI Radar

GPT

Available · Language · Milestone

GPT is a transformer language model from OpenAI, published in June 2018 with a paper and released with its code and pretrained weights under the MIT licence. It was pre-trained without supervision on BooksCorpus and then fine-tuned on smaller supervised datasets for specific tasks; OpenAI released it so others could skip the costly pre-training step. [1] [2] [3] Primary source

Timeline of GPT →

Claims and evidence

  • Released Primary source[1]page date; first paragraph and Drawbacks [3]
  • Status AvailableArchived repository; the pretrained weights (params_0.npy to params_9.npy) remain downloadable from OpenAI on GitHub. Primary source[3]model folder
      • Input text Primary source[2]section 3 Framework
      • Output text Primary source[2]section 3 Framework
      • Open weights Yes Primary source[1]Drawbacks, Compute Requirements [3]model folder
      • Context window 512 tokens tokensTraining sequence length: the paper trains on contiguous sequences of 512 tokens, the post says sequences of up to 512 tokens. Primary source[2]section 4.1 Setup, Model specifications [1]Drawbacks, Compute Requirements
      • Parameters Not disclosedOpenAI describes the architecture (12-layer decoder-only transformer, 768-dimensional states, 12 attention heads) but gives no parameter count in the GPT materials. The GPT-2 paper calls its 117M model equivalent to the original GPT, and the GPT-2 repository later said that count was wrong, so no figure is recorded. No claim madeNo source
      • Access Open-weights download Primary source[3]model folder

      Lineage

      Predecessors

      No known predecessor.

      Successors

      • GPT-2 · 14 February 2019

      Based on

      Not derived from another model.

      Variants and derived

      None recorded.

      Siblings

      None recorded.

      All ancestors

      None.

      All descendants

      Variants

      No variants recorded in this record.

      Related AI Radar coverage

      AI Radar coverage starts in June 2026; no coverage linked yet.

      All model releases from OpenAI on AI Radar →