gpt-oss-120b and gpt-oss-20b
OpenAI's first open-weight language models since GPT-2: Apache 2.0 reasoning models, 120B near o4-mini and 20B near o3-mini.
Mixture-of-experts Transformers with 117B total / 5.1B active and 21B / 3.6B active parameters, 128K context, post-trained like o4-mini with high-compute RL and configurable reasoning effort. 120B fits one 80 GB GPU, 20B runs in 16 GB. Tokenizer o200k_harmony released too.
- Date
- Tuesday, 5 August 2025
- Lab
- OpenAI
- Kind
- open-weights
- Access
- open weights
- Price
- free to download and run
Figures
| Measure | Value | Measured by |
|---|---|---|
| GPQA Diamond, no tools (gpt-oss-120b / 20b) | 80.1% / 71.5% o3 83.3%, o4-mini 81.4%, o3-mini 77% | company |
| Total / active parameters (120b) | 117B / 5.1B 20b: 21B / 3.6B | company |
| Red-teaming challenge prize fund | $500,000 adversarial fine-tuning also tested under Preparedness Framework | company |
Arrived after DeepSeek-R1 and Qwen/Kimi open releases had pulled open-model leadership to China (B20). Independent providers showed uneven quality at launch (Simon Willison, 2025-08-15). A follow-up was gpt-oss-safeguard (2025-10-29).
Sources
- openai.com/index/introducing-gpt-oss/
- simonwillison.net/2025/Aug/5/gpt-oss/
- deploymentsafety.openai.com/gpt-oss/
This record was checked against its sources on 6 October 2026. How we check