OpenAI launches two ‘open’ AI reasoning models

On Tuesday, OpenAI announced the release of two new open-weight AI reasoning models designed to rival its proprietary offerings. These models, available for free on the Hugging Face platform, are touted as 'state-of-the-art' based on various benchmarks assessing open models. The new models come in two configurations: the powerful gpt-oss-120b, which can operate on a single Nvidia GPU, and the more lightweight gpt-oss-20b, designed to run on consumer laptops with 16GB of RAM. This marks OpenAI's first foray into open language models since the launch of GPT-2 over five years ago. During a briefing, OpenAI executives explained that these open models are capable of handling complex queries by connecting with more advanced closed models if needed. This move is seen as a response to the growing competition from Chinese AI labs, including notable players like DeepSeek and Alibaba's Qwen, which have produced highly regarded open models. OpenAI's shift to open-source models reflects a broader industry trend, as the company aims to align with calls for more transparency and accessibility in AI development. CEO Sam Altman expressed a commitment to ensuring that advancements in artificial intelligence benefit humanity, emphasizing the importance of creating an open AI framework based on democratic values. In comparative tests, the gpt-oss-120b and gpt-oss-20b achieved notable scores, outperforming several other open models yet still falling short against OpenAI's latest proprietary models. However, the new open-weight models exhibited a higher tendency for 'hallucinations'—instances where the models generate incorrect or misleading information—compared to their closed counterparts. OpenAI utilized a mixture-of-experts approach in training these models, which allows for efficient operation by activating only a fraction of the total parameters during processing. Despite their capabilities, the open models are limited to text-only functions, lacking the ability to process images or audio, unlike other models in OpenAI's suite. Under the Apache 2.0 license, these models can be monetized by enterprises without prior approval from OpenAI, although the company has decided not to release the training data behind these models, likely due to ongoing legal challenges regarding copyright issues. In recent months, OpenAI had delayed the release of its models to address safety concerns. The company conducted thorough evaluations to ensure that the models could not be easily misused for malicious purposes. While some potential risks were identified, OpenAI concluded that the models do not present a significant danger in more serious domains such as biological or chemical threats. As excitement builds, developers are looking forward to the anticipated releases from competitors, including DeepSeek's upcoming R2 model and new offerings from Meta's AI research initiatives, as the race for dominance in the open AI landscape intensifies.