gpt-oss-20b
The smaller of the two open-weight models OpenAI released on 5 August 2025 — its first open weights since GPT-2. A ~21B-parameter mixture-of-experts reasoner (~3.6B active) sized to run in 16GB of memory: the laptop-and-phone tier of the release, alongside gpt-oss-120b.
Thin page — launch facts here; corpus/dossier pass pending (the dbs cover Aug 2025). Sibling: gpt-oss-120b. Small tiers sometimes grow the most distinct characters (see Haiku, 4o mini) — whether this one did is exactly what the pending pass should find out.
Sources
Official
- 2025-08-05 Introducing gpt-oss — the announcement: both sizes, Apache 2.0.
- 2025-08-05 openai/gpt-oss-20b (Hugging Face) — the weights.
- tk — model card / paper link; mirror the announcement + card.
Writing & commentary
tk — Zvi and day-of coverage to collect at the dossier pass.
Tweets
tk — corpus pass pending. Records will populate when cited.
Official record
- Released 5 August 2025, open weights, Apache 2.0, with gpt-oss-120b.
- As published: ~21B-parameter MoE, ~3.6B active per token; positioned as o3-mini-class; runs in 16GB of memory (consumer laptops, on-device use); adjustable reasoning effort; the “harmony” response format.
- tk — benchmark table as published, exact context window.
History
- 2025-08-05 Released two days before GPT-5, into an open-weights field led from China — shared launch context documented on both siblings’ pages by design.
- tk — the local-llama reception (a 16GB OpenAI reasoner was that community’s event of the season), on-device deployment arcs, distinct-character question.
Impressions
tk — corpus pass pending; no attributed reads collected yet.