OpenAI has released two open-weight language models—gpt-oss-120b and gpt-oss-20b—marking its first publicly downloadable large-language models since GPT-2 in 2019. The company says the larger, 117-billion-parameter model achieves near-parity with its proprietary o4-mini system on core reasoning benchmarks, while the 21-billion-parameter version matches o3-mini. Both models use a Mixture-of-Experts design that activates only a fraction of their total parameters, allowing gpt-oss-120b to run on a single 80 GB Nvidia H100 GPU and gpt-oss-20b to operate on consumer hardware with 16 GB of video memory. They support context windows of up to 128 000 tokens and expose full chain-of-thought reasoning for inspection and debugging. The weights are released under the permissive Apache 2.0 license, enabling commercial use, fine-tuning and redistribution. OpenAI subjected the systems to additional adversarial testing and is offering a US$500,000 red-teaming programme after external reviewers found the models did not reach high-risk capability thresholds. Downloads are already live on platforms including Hugging Face, Azure, Groq and Together AI. Amazon Web Services said the models are available in Bedrock and SageMaker, claiming cost-performance advantages of up to three times over Alphabet’s Gemini and twice over OpenAI’s own o4 models when run in its cloud. The move returns OpenAI to its open-source roots and intensifies competition with Meta’s Llama series, DeepSeek-R, Mistral and other open models. It also arrives amid expectations that OpenAI will unveil GPT-5 later this year, underscoring the rapid escalation in large-model releases and partnerships across the generative-AI industry.