OpenAI returns to open-source with first free models in six years

OpenAI released its first open-weight models since 2019 on Tuesday, directly challenging Chinese companies that now dominate the accessible artificial intelligence (AI) space with powerful, free-to-use technology.

The two new large language models (LLMs), gpt-oss-120b and gpt-oss-20b, can run on personal computers and laptops. They perform nearly as well as OpenAI’s paid reasoning models in coding and math tasks.

“Going back to when we started in 2015, OpenAI’s mission is to ensure AGI (Artificial General Intelligence) that benefits all of humanity,” CEO Sam Altman said in a statement. “We are excited for the world to be building on an open AI stack created in the United States, based on democratic values.”

The move puts OpenAI back in competition with Chinese firms like DeepSeek and Alibaba’s Qwen series. These companies shocked the tech world earlier this year by releasing powerful AI models that anyone could download and use for free.

OpenAI had avoided open releases for six years, focusing instead on selling access to its models through ChatGPT and business services. Critics started calling the company “ClosedAI” for abandoning its original mission of making AI widely available.

The smaller model, gpt-oss-20b, needs just 16GB of computer memory to run. The larger version requires specialized graphics cards with 80GB of memory.

“Open-weight models have a very different set of strengths,” said OpenAI co-founder Greg Brockman. “People can run them behind their firewall.”

Both models use chain-of-thought reasoning, taking extra time to work through complex problems step by step. They can browse websites, write code, and call other AI tools when needed.

OpenAI tested the models on competitive programming challenges. The larger model scored 2622 points while the smaller earned 2516 points. Both outperformed DeepSeek’s latest model, which scored 2463 points.

However, the new models only work with text, unlike OpenAI’s main services, which can handle images and voice. More concerning, they hallucinate frequently, with error rates between 49-53% compared to just 16% for OpenAI’s premium o1 model. Users wanting those features must still pay for cloud-based access.

The release follows President Trump’s AI Action Plan, which encouraged American companies to develop open models as “global standards” based on democratic values.

Meta previously led the American open AI movement with its Llama models. But the company now focuses on building superintelligent AI and may stop releasing open models due to safety worries.

OpenAI delayed these models twice for safety testing. They conducted extensive safety testing, even trying to fine-tune the models for malicious purposes. The company says the models never reached dangerous capability levels.

Both models are available under Apache 2.0 licensing, allowing companies to use them commercially without paying fees. Amazon announced the models would appear on its Bedrock AI marketplace.

Companies including Orange SA, Snowflake, and AI Sweden helped test the models before release. OpenAI plans to use feedback from users to decide whether to release more open models in the future.