OpenAI Releases Open-Weight Models Amid Rising Open-Source Competition

OpenAI has launched its first open-weight model in over four years, marking a significant shift in the company’s strategy amid rising competition from open-source alternatives. The two new models, gpt-oss-120b and gpt-oss-20b, are designed to advance the frontier of reasoning in open-source AI while maintaining the proprietary elements that underpin more advanced systems like GPT-5. CEO Sam Altman described the release as an effort to “get AI into the hands of the most people possible,” emphasizing the company’s commitment to democratizing access to AI innovation [1].

The release comes after Altman acknowledged earlier this year that OpenAI had been “on the wrong side of history” in its reluctance to open up its models. This admission followed the success of Chinese open-source models such as DeepSeek, which demonstrated strong reasoning capabilities at a lower cost. OpenAI’s new models are tailored for developers and AI builders, with the larger model capable of running on a single Nvidia 80GB chip and the smaller model suitable for consumer hardware like Mac laptops. This positioning suggests the models are intended to be used in practical, real-world applications rather than as fully open tools [1].

Despite the open-weight designation, the models are not fully transparent. They use a Mixture-of-Experts (MoE) architecture, but only the publicly known components are included. Critical elements such as the routing mechanisms and training methods remain confidential. Dylan Patel of SemiAnalysis noted that OpenAI’s approach is deliberate—by using only widely known building blocks, the company avoids leaking valuable intellectual property while still providing a useful model to the public [1].

Greg Brockman, co-founder and president of OpenAI, stated that the new models are complementary to the company’s existing offerings and align with its mission to ensure AI benefits all of humanity. However, OpenAI declined to benchmark the new models against Chinese open-source systems like DeepSeek or Qwen, despite reports that these models have recently outperformed U.S. competitors on key reasoning benchmarks. The company said it is confident in its internal benchmarks and left it to the broader AI community to conduct further comparisons [1].

Aleksa Gordic, a former Google DeepMind researcher, explained that this strategy—sharing a useful but limited model—allows companies like OpenAI to support developers and startups while protecting their core intellectual property. He noted that Meta and Mistral have adopted similar approaches, releasing open models without exposing their most valuable innovations [1].

OpenAI’s decision signals a strategic response to the growing influence of open-source AI. While the company remains focused on maintaining its proprietary edge, it is now engaging more directly with the broader AI ecosystem. As the landscape continues to evolve, the balance between open access and proprietary innovation will remain a defining challenge in the development of artificial intelligence [1].

Source: [1] OpenAI launches its first open model in years so it can stop being on the ‘wrong side of history’—while still keeping its most valuable IP under wraps (https://fortune.com/2025/08/05/openai-launches-open-source-llm-ai-model-gpt-oss-120b-deepseek/)