GPT-2

Available · Language · Milestone

GPT-2 is a transformer language model from OpenAI with up to 1.5 billion parameters, trained on about 40GB of internet text and described by OpenAI as a direct scale-up of GPT. Citing concerns about malicious use, OpenAI released its weights in stages during 2019, from a 124M model in February to the 1.5B model in November. [1] [2] [3] Primary source

Timeline of GPT-2 →

Claims and evidence

  • Released Staged release: on 2019-02-14 OpenAI published the paper and only a much smaller model; the full 1.5B model followed on 2019-11-05 (see variants). Primary source[1]page date; second paragraph [2]first paragraph
  • Status AvailableArchived repository; the weights of all four sizes remain downloadable from OpenAI storage through the repository download script. Primary source[4]download_model.py [3]
  • Successor of GPTOpenAI calls GPT-2 a successor to GPT and a direct scale-up of it. Primary source[1]second and third paragraphs
  • Change · Size More than ten times the parameters of GPT, with 1.5 billion in the biggest version.Compared with GPT Primary source[1]
  • Change · Training data Trained on WebText, about 40GB of text from 8 million web pages, more than ten times the training data of GPT.Compared with GPT Primary source[1] [5]
  • Change · Context length Context size doubled from 512 to 1024 tokens.Compared with GPT Primary source[6]
  • Change · Availability Weights released in stages over nine months, starting with the smallest model, instead of all at once.Compared with GPT Primary source[2] [3]
  • Input text Primary source[5]Model type [6]section 2
  • Output text Primary source[1]Samples [6]section 2
  • Open weights YesWeights released in stages between February and November 2019 under a Modified MIT licence. Primary source[3]first paragraph [4]download_model.py; License
  • Context window 1024 tokens tokens Primary source[6]section 2.3 Model
  • Parameters 1.5 billionThe largest size; the paper lists it as 1542M and the 6-month follow-up as 1558M. Smaller sizes are in variants. Primary source[1]third paragraph [5]Model version
  • Access Open-weights download Primary source[4]download_model.py [3]

Lineage

Predecessors

  • GPT · 11 June 2018

Successors

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

  • GPT · 11 June 2018

All descendants

Variants

GPT-2 124M

  • Released Primary source[1]second paragraph [2]first paragraph

Inline variant in this record. The much smaller model released with the announcement.

Also known as: GPT-2 117M, GPT-2 small

Differs in:

  • Parameters: 124M Evidence not assessed [2]first paragraph [5]Model version

GPT-2 355M

  • Released Primary source[1]GPT-2 Interim Update, May 2019 [2]first paragraph

Inline variant in this record.

Also known as: GPT-2 345M, GPT-2 medium

Differs in:

  • Parameters: 355M Evidence not assessed [2]first paragraph [5]Model version

GPT-2 774M

  • Released Primary source[2]page date; first paragraph

Inline variant in this record.

Also known as: GPT-2 large

Differs in:

  • Parameters: 774 million Evidence not assessed [2]first paragraph [5]Model version

GPT-2 1.5B

  • Released Primary source[3]page date; first paragraph

Inline variant in this record. The 1.5B-parameter version, which the paper and the announcement call GPT-2; the final step of the staged release, published with code and model weights.

Also known as: GPT-2 XL

Related AI Radar coverage

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

All model releases from OpenAI on AI Radar →