AI Radar

OPT

Available · Language · Milestone

OPT (Open Pre-trained Transformers) is a suite of decoder-only language models from Meta, ranging from 125 million to 175 billion parameters, released on 3 May 2022 together with the metaseq code used to train them. The sizes up to 66B were released for download under a non-commercial licence; OPT-175B went to approved researchers on request. [1] [2] [3]README: Pretrained Model Weights Primary source

Timeline of OPT →

Claims and evidence

  • Released Primary source[2]page date: May 3, 2022 [1]Appendix D, Model Card: Model date
  • Status AvailableWeights of OPT-125M to OPT-66B are downloadable from Meta's facebook organisation on Hugging Face and through the links in the metaseq README (seen 2026-10-01). The metaseq repository itself is archived (read-only). OPT-175B is available on request only; see that variant. Primary source[4] [3]README: Pretrained Model Weights
      • Open weights YesApplies to OPT-125M to OPT-66B, released for download under the non-commercial OPT-175B License Agreement. OPT-175B weights are shared with approved researchers on request. Primary source[2]Responsible publication with OPT-175B [1]Section 1 Introduction [3]README: Pretrained Model Weights
      • Context window 2048 tokensThe paper states that all OPT models were trained with a sequence length of 2048; it gives no separate context-window figure. Primary source[1]Section 2.2 Training Setup
      • Access Open-weights downloadOPT-175B differs; see that variant. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Lineage

      Predecessors

      No known predecessor.

      Successors

      No known successor.

      Based on

      Not derived from another model.

      Variants and derived

      None recorded.

      Siblings

      None recorded.

      All ancestors

      None.

      All descendants

      None.

      Variants

      OPT-125M

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-125m

      Differs in:

      • Parameters: 125M Evidence not assessed [3]README: Pretrained Model Weights

      OPT-350M

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-350m

      Differs in:

      • Parameters: 350M Evidence not assessed [3]README: Pretrained Model Weights

      OPT-1.3B

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-1.3b

      Differs in:

      • Parameters: 1.3B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-2.7B

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-2.7b

      Differs in:

      • Parameters: 2.7B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-6.7B

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-6.7b

      Differs in:

      • Parameters: 6.7B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-13B

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-13b

      Differs in:

      • Parameters: 13B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-30B

      Same dates as OPT.

      • Variant Inline variant in this record. Primary source[2]Responsible publication with OPT-175B [3]README: Pretrained Model Weights

      Also known as: facebook/opt-30b

      Differs in:

      • Parameters: 30B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-66B

      • Announced The launch post listed the 66B model as to be released soon. Primary source[2]Responsible publication with OPT-175B
      • Released Day the pull request adding the OPT-66B download links was merged in Meta's metaseq repository (UTC). The facebook/opt-66b repo on Hugging Face was created on 2022-06-23 according to its repository metadata, which is not a stated date. Primary source[5]
      • Variant Inline variant in this record. Primary source[5] [3]README: Pretrained Model Weights

      Also known as: facebook/opt-66b

      Differs in:

      • Parameters: 66B Evidence not assessed [3]README: Pretrained Model Weights

      OPT-175B

      Same dates as OPT.

      • Variant Inline variant in this record.Not on Hugging Face; Meta grants access to academic researchers, people at government, civil society and academic organisations, and industry research labs. Primary source[2] [1]Section 1 Introduction

      Status: Research only Evidence not assessed [2]opening section [3]README: Pretrained Model Weights, request access

      Also known as: OPT 175B

      Differs in:

      • Parameters: 175B Evidence not assessed [3]README: Pretrained Model Weights [2]opening section
      • Access: research only Evidence not assessed [2]opening section [1]Section 1 Introduction

      Related AI Radar coverage

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

      All model releases from Meta on AI Radar →