OpenAI launches two ‘open’ AI reasoning models

On Tuesday, OpenAi announced the launch of two Open-Wight AI-reasoning models with similar possibilities such as his O-series. Both are freely available for downloading from the online developer platform, HugThe company said and described the models as “state-of-the-art” when measured in different benchmarks for comparing open models.

The models come in two sizes: a larger and more capable GPT-OSS-120B model that can be performed on a single NVIDIA GPU and a lighter-weight GPT-OSS-20B model that can be performed on a consumer laptop with 16 GB of memory.

The launch marks the first ‘open’ language model of OpenAi since GPT-2, which was released more than five years ago.

In a briefing, OpenAi said that its open models will be able to send complex queries to AI models in the cloud, as WAN reported earlier. This means that if the open model of OpenAi is unable to do a certain task, such as processing an image, developers can connect the open model to one of the more capable closed models of the company.

While OpenAi opened AI models open-sourced AI models in its beginning, the company has generally preferred its own closed-source development approach. The last strategy has helped to build a large company that sells access to its AI models through an API of companies and developers.

CEO Sam Altman, however, said in January that he believes that OpenAi has been “on the wrong side of history” when it comes to open purchasing of his technologies. Today, the company is confronted with growing pressure from the Chinese AI Laboratories – including Deepseek, Alibaba’s Qwen and Moonshot AI – that have developed several of the world’s most capable and popular open models. (While Meta previously dominated the Open AI room, the Llama AI models of the company were left in the past year.)

In July, the Trump administration also insisted on American AI developers to open more technology to promote the global acceptance of AI that are tailored to American values.

With the release of GPT-Oss, OpenAI is hoping for both developers and the Trump administration, both of which have seen the Chinese AI Laboratories rise to fame in the open source room.

“Going back to when we started in 2015, the mission of OpenAI is to ensure that Agi that benefits all humanity,” said OpenAi CEO Sam Altman in a statement shared with WAN. “To that end, we are enthusiastic that the world is building on an open AI stack made in the United States, based on democratic values, available to everyone for free and for a broad advantage.”

How the models performed

OpenAi wanted to make its open model a leader, among other things, open-weight AI models, and the company claims to have done exactly.

On code forces (with tools), a competitive coding test, GPT-OSS-120B and GPT-OSS-20B score 2622 and 2516, respectively, the Deepseek’s R1, respectively, surpassed with O3 and O4-Mini.

At the last exam of humanity, a challenging test of crowd-sourced questions on various topics (with tools), GPT-OSS-13B and GPT-OSS-20B, 19% and 17.3% respectively. Likewise, this interpreted O3 but performs better than leading open models of Deepseek and QWen.

In particular, the open models of OpenAI hallucinate considerably more than the latest AI-reasoning models, O3 and O4-Mini.

Hallucinations have become more serious in the newest AI Reasoning models from OpenAi, and the company said rather that it does not fully understand why. In a white paper, OpenAI says that this is “expected, because smaller models have less world knowledge than larger border models and tend to hallucinate more.”

OpenAI discovered that GPT-OSS-1320B and GPT-OSS-20B hallucinated in response to 49% and 53% of the questions about Personqa, the company’s internal benchmark for measuring the accuracy of the knowledge of a model about people. That is more than triple the hallucination percentage of the O1 model of OpenAi, which scored 16%, and higher than the O4-mini model, which scored 36%.

Training of the new models

OpenAi says that his open models are trained with similar processes such as his own models. The company says that every open model uses mixture of experts (tired) to tap fewer parameters for a certain question, making it more efficient. For GPT-OSS-120B, which has 117 billion total parameters, OpenAI says that the model activates only 5.1 billion parameters per token.

The company also says that its open model has been trained with the help of high-compute Reinforcement Learning (RL)-a process after training to learn AI models right of incorrectly in simulated environments using large clusters of Nvidia GPUs. This was also used to train OpenAi’s O-series models, and the open models have a similar process of thought process in which they take extra time and computational sources to go through their answers.

As a result of the process after the training, OpenAI says that the open AI models excel in feeding AI agents, and are able to call tools such as web search or python code version as part of the process of the chain of the debating. However, OpenAi says that the open models are only text, which means that they cannot process or generate images and audio, such as the other models of the company.

OpenAi releases GPT-OSS-120B and GPT-OSS-20B under the Apache 2.0 license, which is generally considered one of the most tolerant. With this license, Enterprises can earn the Open models of OpenAI without having to pay or receive permission from the company.

In contrast to fully open source offers from AI Labs such as AI2, Openai says that it will not release the training data that is used to make its open models. This decision is not surprising, since various active lawsuits against AI model providers, including OpenAI, have claimed that these companies have wrongly trained their AI models on copyrighted works.

OpenAI has delayed the release of its open models several times in recent months, partly to tackle safety problems. In addition to the typical safety policy of the company, OpenAi says in a white paper that it also investigated whether bad actors can refine GPT-OSS models to be more useful in cyber attacks or creating organic or chemical weapons.

After testing OpenAi and external evaluators, the company says that GPT-Oss can increase the biological possibilities marginally. However, it found no evidence that these open models could reach the threshold of ‘high capacities’ for danger in these domains, even after coordination.

Although the model of OpenAI State-of-the-Art seems to be among open models, developers are eagerly waiting for the release of Deepseek R2, the next AI-reasoning model, as well as a new open model of the new superintelligence lab from Meta.