OpenAI has launched two open-weight AI reasoning models, gpt-oss-120b and gpt-oss-20b, marking its first “open” release since GPT-2. Available on Hugging Face under the Apache 2.0 license, the models target developers seeking high-performance AI tools without restrictive licensing. The larger 120b model runs on a single Nvidia GPU, while the smaller 20b version can operate on consumer laptops with 16GB RAM.
The models are designed to power AI agents, supporting tool use such as web search or Python execution. They can route complex queries to OpenAI’s proprietary models when needed. Benchmarks show gpt-oss models outperform leading open competitors like DeepSeek R1 and Qwen in coding and general reasoning tasks, though they underperform OpenAI’s closed o-series models.
Key performance notes:
- Codeforces (with tools): 120b scored 2622; 20b scored 2516.
- Humanity’s Last Exam: 120b scored 19%, 20b scored 17.3%.
- Higher hallucination rates than o-series models, with 49–53% on PersonQA, versus 16% for o1.
Technically, the models employ a mixture-of-experts (MoE) architecture, activating just 5.1B of 117B parameters per token for efficiency. They were trained with high-compute reinforcement learning (RL) and a chain-of-thought approach similar to OpenAI’s proprietary models, but remain text-only, without image or audio processing capabilities.
OpenAI withheld training data to mitigate copyright concerns but confirmed safety testing against misuse in cyber, chemical, or biological domains, finding no “high capability” risk. The release reflects CEO Sam Altman’s pivot toward open AI development amid competition from Chinese labs like DeepSeek and Qwen, and U.S. government encouragement for open AI aligned with democratic values.
With DeepSeek R2 and Meta’s upcoming open models on the horizon, the launch positions OpenAI to regain influence in the rapidly evolving open AI agent ecosystem.
Source:
https://techcrunch.com/2025/08/05/openai-launches-two-open-ai-reasoning-models/
