Hello, tech lovers! Today, we’re diving into some exciting news from the AI world, so buckle up!

OpenAI has just launched two new open-weight AI reasoning models that match the capabilities of its previous o-series models. These models are freely downloadable from Hugging Face and are praised for their top-tier performance across various benchmarks. The big guys come in two sizes: a powerful gpt-oss-120b that runs smoothly on a single Nvidia GPU, and a lighter gpt-oss-20b that can work on a regular laptop with 16GB of RAM.

This marks OpenAI’s first open-source model since GPT-2, released over five years ago. The company explained that the new models can communicate with more advanced AI cloud models, making them more versatile. For instance, if the open model can’t handle a task like image processing, it can connect with a more robust closed model for help. Historically, OpenAI leaned toward keeping models proprietary, mainly to build a commercial business through API access. However, CEO Sam Altman mentioned they now see open sourcing as the right path, especially with the rise of Chinese AI labs like DeepSeek and Qwen, and with U.S. officials advocating for more open AI tech.

The models are under the Apache 2.0 license, enabling enterprises to monetize them freely. Yet, OpenAI won’t release the training data, citing ongoing legal issues around copyrighted material used in training. They have also conducted safety checks to prevent misuse, including assessing the models’ potential for harmful applications, which they found to be marginally increased but not beyond safety thresholds. Performance-wise, the models excel in coding tests and other benchmarks, though they tend to hallucinate more than larger, proprietary models.

Training these models involved advanced methods like mixture-of-experts to improve efficiency, with the gpt-oss-120b activating only a small part of its parameters per query. Reinforcement learning was used post-training to align the models with ethical standards, and chain-of-thought techniques allow them to incorporate tool use like web searches or coding. Still, these open models are text-only and don’t handle images or audio. OpenAI hopes these models will inspire the community and help regain favor among developers and policymakers alike.

As the AI race heats up, all eyes are on upcoming models from competitors like DeepSeek R2 and Meta. OpenAI’s move is seen as a strategic step to foster an open, democratic AI ecosystem and counter the growth of Chinese labs dominating the open model space.