OpenAI has further launched officially two new open-weight AI reasoning models— gpt-oss-120b, gpt-oss-20b. With these releases, the company hopes to enter the open-weight AI solutions leader’s circle. These models are primarily concerned with enhancing capacity. They address some of the safety issues that have held them up the past several months. OpenAI claims that the gpt-oss-120b model is “best in class” among its peers, showcasing a commitment to innovation amid growing competition from Chinese AI labs such as DeepSeek, Alibaba’s Qwen, and Moonshot AI.
Both models are based on a mixture-of-experts (MoE) architecture. This arrangement allows them to use a smaller number of active parameters per query, improving their effectiveness. The gpt-oss-120b has been specifically designed to utilize only 117 billion total parameters while activating only 5.1 billion parameters per token. Of all these models, the gpt-oss-20b model is most notable for its accessibility. In fact, it can run quite smoothly on a consumer laptop with only 16GB of memory.
Performance Insights
OpenAI has positively assessed the models in terms of performance metrics. Specifically, they zeroed in on how frequently the models produce false information, a behavior colloquially dubbed as “hallucination.” The other company ultimately determined that the gpt-oss-120b halluciated 49% of the answers on the PersonQA dataset. In comparison, the gpt-oss-20b did marginally worse at a 53% hallucination rate.
OpenAI admits that these findings are “unsurprising.” So world knowledge diminishes proportionally with model size. Smaller models will thus often be less knowledgeable than larger frontier models who hallucinate less. This emphasizes the immense difficulty that remains in achieving any kind of accuracy, any kind of reliability in AI models, even ones with many more parameters.
“expected, as smaller models have less world knowledge than larger frontier models and tend to hallucinate more.” – OpenAI
Strategic Objectives
OpenAI’s decision to release both models under the Apache 2.0 license aligns with its strategic objectives to promote transparency and collaboration within the AI research community. That was CEO Sam Altman’s stated goal when he founded OpenAI – to make sure that widespread developments in artificial intelligence brought benefits to all humans.
“Going back to when we started in 2015, OpenAI’s mission is to ensure AGI that benefits all of humanity,” – Sam Altman
As AI technology advances, so too do OpenAI’s trickery and techniques. We are endeavoring to raise the bar on what open-source tools can be. Their skin is almost as thick as the competition. The disorganization gets the shark tank treatment. It’s doing a lot to remain a leader in AI research and development.
Addressing Safety Concerns
This delay in releasing these models is not without merit, as we know that OpenAI put safety first. To that end, they prioritized a deployment that was responsible. The organization understands that using AI responsibly is critical in any field, but especially in a field where misuse can have deadly outcomes. By focusing on getting these models right over time, OpenAI aims to minimize potential risks and build user confidence.
OpenAI’s proactive approach reflects a broader trend within the tech industry, where companies are increasingly prioritizing ethical considerations alongside technological advancements. The emphasis on safety and reliability is interpreted as a key part of building public trust in AI technologies.