OpenAI has released its first open-weight language models since 2019, offering developers two versions—gpt-oss-120b and the lighter gpt-oss-20b—under the permissive Apache 2.0 licence. The larger model contains roughly 117 billion parameters but activates just 5.1 billion per token, allowing it to run on a single 80-GB GPU, while the 20-billion-parameter variant can operate on a consumer laptop with 16 GB of memory. Both models are optimised for advanced reasoning, coding and maths tasks and perform on par with the company’s proprietary o3-mini and o4-mini systems. The weights can be downloaded via Hugging Face and are being distributed through Amazon Web Services’ Bedrock marketplace, the first time an OpenAI model has appeared on the platform. Telecoms group Orange and cloud-data specialist Snowflake are among early testers, and additional hosting is planned on Groq’s inference cloud and Saudi AI provider Humain. OpenAI says the move responds to developer demand and rising competition from Chinese open-weight rivals such as DeepSeek R1 and Alibaba’s Qwen, while aligning with Washington’s call for more transparent AI stacks built in the United States. The company has also launched a US$500,000 red-teaming challenge, but hackers claimed to have bypassed the models’ safety guard-rails within hours of the release. Separately, people familiar with the matter say OpenAI is organising a secondary share sale that would let staff cash out and could lift the Microsoft-backed firm’s valuation to about US$500 billion, up from US$300 billion in April. CNBC reported that the tender could exceed US$1.5 billion. The product launch comes ahead of a Thursday livestream scheduled for 10 a.m. Pacific Time, widely expected to feature further model announcements after chief executive Sam Altman hinted at the imminent debut of GPT-5.