OpenAI launches new gpt-oss models optimized for local use, excelling in coding and health queries

OpenAI said it has released two open-weight language models that excel in advanced reasoning and are optimized to run on laptops with performance levels similar to its smaller proprietary reasoning models. An open-weight language model’s trained parameters or weights are publicly accessible, which can be used by developers to analyze and fine-tune the model for specific tasks without requiring original training data. 

“One of the things that is unique about open models is that people can run them locally. People can run them behind their own firewall, on their own infrastructure,” OpenAI co-founder Greg Brockman said in a press briefing. Open-weight language models are different from open-source models, which provide access to the complete source code, training data and methodologies. 

Separately, Amazon announced OpenAI’s open-weight models are now available on its Bedrock generative AI marketplace in Amazon Web Services. It marks the first time an OpenAI model has been offered on Bedrock, said Atul Deo, Bedrock’s director of product. “OpenAI has been developing great models and we believe that these models are going to be great open-source options, or open-weight model options for customers,” Reuters reported, citing Deo.  

Amazon shares tumbled last week after the company reported slowing growth in its AWS unit, particularly compared with rivals. The landscape of open-weight and open-source AI models has been highly contested this year. 

For a time, Meta’s Llama models were considered the best, but that changed earlier this year when China’s DeepSeek released a powerful and cost-effective reasoning model, while Meta struggled to deliver Llama 4. The two new OpenAI models are the first open models OpenAI has released since GPT-2, which was released in 2019. 

Performance capabilities of OpenAI’s new models

OpenAI’s larger model, gpt-oss-120b, can run on a single GPU, and the second, gpt-oss-20b, is small enough to run directly on a personal computer, the company said. OpenAI said the models have similar performance to its proprietary reasoning models called o3-mini and o4-mini, and especially excel at coding, competition math and health-related queries. The models were trained on a text-only dataset which in addition to general knowledge, focused on science, math and coding knowledge. OpenAI did not release benchmarks comparing the open-weight models to competitors’ models such as the DeepSeek-R1 model. Microsoft-backed OpenAI, currently valued at $300 billion, is currently raising up to $40 billion in a new funding round led by Softbank Group.

Read more: OpenAI to acquire Apple designer’s startup for $6.5 billion to create new AI-powered computers

Enhance accessibility and customization

The newly released open-weight models, gpt-oss-120b and gpt-oss-20b, represent a strategic move by OpenAI to provide accessible and affordable AI tools for developers and researchers, enabling broader customization and deployment flexibility compared to proprietary models. According to OpenAI’s official release, these models underwent advanced post-training techniques, including supervised fine-tuning and reinforcement learning stages identical to those used for proprietary models, enhancing their reasoning and tool-use capabilities. 

Outperforming proprietary models in key benchmarks

Evaluations show that gpt-oss-120b outperforms OpenAI’s own o3-mini and matches or exceeds the o4-mini model on benchmarks encompassing coding competitions (such as Codeforces), general problem-solving (MMLU and HLE), and tool-calling tasks (TauBench). Notably, it surpasses o4-mini on health-related queries (HealthBench) and mathematical competitions (AIME 2024 & 2025), while the smaller gpt-oss-20b model matches or even exceeds the o3-mini on these evaluations despite its reduced size. For performance optimization across hardware, OpenAI partnered with Nvidia, AMD, and Groq to ensure compatibility with diverse GPU architectures. Nvidia CEO Jensen Huang highlighted the significance of these developments in promoting innovation through open-source software collaborations.

Expanding AI resources with OpenAI models

AWS’s integration of OpenAI’s open-weight models into the Bedrock and SageMaker platforms marks a landmark expansion in cloud-accessible AI resources. Amazon claims that the larger gpt-oss-120b model offers a highly competitive price-performance ratio that is reportedly 10 times better than Google’s Gemini model, 18 times better than the DeepSeek-R1 model, and 7 times better than OpenAI’s own proprietary o4 model. This competitive advantage is expected to empower the millions of AWS customers with more cost-effective and powerful AI tools, facilitating use cases across agentic workflows, scientific analysis, mathematical problem-solving, and coding. This move also symbolizes an unprecedented step in OpenAI’s commercial relationships, as these models are made available under an Apache 2.0 license, allowing any cloud provider to host them, thus partially sidestepping Microsoft’s exclusive Azure API rights.

Fierce competition in the open-weight AI model space

The open-weight AI model space remains fiercely competitive globally. In addition to Meta’s Llama and China’s DeepSeek initiatives, Microsoft-backed startup Mistral AI also released open-weight models recently. 

OpenAI’s release renews its leadership in the field after a hiatus since GPT-2. Market analysts anticipate that these developments will accelerate open-source AI adoption and innovation, while reshaping partnerships between major cloud providers and AI companies. The evolving dynamics are crucial as OpenAI simultaneously pursues a substantial $40 billion funding round led by Softbank Group, underscoring the intensifying capital inflow and strategic positioning in the generative AI market.