AI Radar

gpt-oss

Available · Language, Reasoning · Milestone

gpt-oss is a pair of open-weight language models from OpenAI, gpt-oss-120b and gpt-oss-20b, released on 5 August 2025 under the Apache 2.0 licence with reasoning and tool-use abilities. OpenAI calls them its first open-weight language models since GPT-2; both are text-only mixture-of-experts models, with 117B and 21B total parameters. [1] [2] [3] [4] Primary source

Timeline of gpt-oss →

Claims and evidence

  • Released The Hugging Face repositories were created on 2025-08-04 (UTC); OpenAI dates the release and the model card August 5, 2025. Primary source[1]page date; introduction [2]cover date
  • Status AvailableWeights of both models downloadable from OpenAI's Hugging Face organisation when read on 2026-10-01. Primary source[3] [4]
      • Input text Primary source[2]section 1 [5]section 1 [6]Model details
      • Output text Primary source[6]Model details
      • Open weights YesApache 2.0 licence, with the gpt-oss usage policy. Primary source[1]Availability [2]section 1 [3] [4]
      • Context window 131,072 tokensThe announcement gives a native context length of up to 128k for both models. Primary source[2]section 2.2 [6]Model details [7]Model details
      • Access Open-weights download, cloud partnerWeights on Hugging Face; OpenAI lists deployment partners such as Azure and AWS. The announcement says OpenAI may consider API support later. Primary source[1]Availability

      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

      gpt-oss-120b

      Same dates as gpt-oss.

      • Variant Inline variant in this record.36 layers, 128 experts with 4 active per token; runs on a single 80 GB GPU with MXFP4-quantised MoE weights. Primary source[1]introduction [3] [2]section 2

      Also known as: gpt-oss 120b

      Differs in:

      • Parameters: 117B total, 5.1B active per token Evidence not assessed [1]Pre-training & model architecture [6]introduction

      gpt-oss-20b

      Same dates as gpt-oss.

      • Variant Inline variant in this record.24 layers, 32 experts with 4 active per token; runs within 16 GB of memory. Primary source[1]introduction [4] [2]section 2

      Also known as: gpt-oss 20b

      Differs in:

      • Parameters: 21B total, 3.6B active per token Evidence not assessed [1]Pre-training & model architecture [7]introduction

      Related AI Radar coverage

      All model releases from OpenAI on AI Radar →