Two “Open” AI Reasoning Models Released by OpenAI

OpenAI competes with Mark Zuckerberg’s Meta by releasing its own open-source AI models.

The creator of ChatGPT has revealed two open-weight large language models that users may tweak and download for free.

These two open-weight AI reasoning models that are comparable to OpenAI’s o-series were released earlier today, Tuesday. Hugging Face, an online developer platform, said that both are publicly available for download. The business described the models as “state of the art” when compared to a number of benchmarks for open models.

Similar terms apply to Meta’s Llama models, and OpenAI’s move differs from ChatGPT, which is predicated on a “closed” model that isn’t customisable.

The models are available in two sizes: the compact gpt-oss-20b model, which can run on a consumer laptop with 16GB of RAM, and the larger and more powerful gpt-oss-120b model, which can run on a single Nvidia GPU.

With this launch, OpenAI has released its first “open” language model since GPT-2, which came out over five years ago.

A media agency which had previously reported, OpenAI stated in a briefing that its open models will be able to transmit sophisticated queries to cloud-based AI models. This implies that developers can link OpenAI’s open model to one of the company’s more powerful closed models if the open model is unable to do a particular task, like processing an image.

Although OpenAI initially made its AI models publicly available, the company has often preferred a closed-source, proprietary development methodology. The latter tactic has assisted OpenAI in becoming a sizable business by granting developers and businesses access to its AI models through an API.

On the other hand, OpenAI has been “on the wrong side of history” in terms of open sourcing its technologies, according to CEO Sam Altman’s January statement. Chinese AI labs, such as DeepSeek, Alibaba’s Qwen, and Moonshot AI, have created some of the most powerful and well-liked open models in the world, and they are putting increasing pressure on the firm today. (Meta used to be the industry leader in open AI, but in the past year, their Llama AI models have lagged.)

In order to encourage the worldwide use of AI that is consistent with American principles, the Trump administration also asked American AI developers to open source more of their technology in July.

Having seen the Chinese AI laboratories gain notoriety in the open source domain, OpenAI intends to win over developers and the Trump administration with the release of gpt-oss.

“OpenAI’s mission is to ensure AGI that benefits all of humanity, going back to when we started in 2015,” Altman told TechCrunch in a statement. In light of this, we are thrilled that the world will be able to build upon an open AI stack developed in the US, founded on democratic principles, and made freely and widely available to everyone.

Based on the models’ performance, the company says it has achieved its goal of making its open model the best among other open-weight AI models.

The gpt-oss-120b and gpt-oss-20b scores on Codeforces (with tools), a competitive coding test, are 2622 and 2516, respectively, exceeding DeepSeek’s R1 and underperforming o3 and o4-mini.

Gpt-oss-120b and gpt-oss-20b score 19% and 17.3%, respectively, on Humanity’s Last Exam (HLE), a difficult test consisting of crowdsourced questions in a range of areas (with tools). In a similar vein, this performs better than top open models from DeepSeek and Qwen but worse than o3. 

Notably, compared to its most recent AI reasoning models, o3 and o4-mini, OpenAI’s open models exhibit noticeably more hallucinations.

OpenAI previously stated that it doesn’t fully understand why hallucinations have been becoming more severe in its most recent AI reasoning models. This is “expected, as smaller models have less world knowledge than larger frontier models and tend to hallucinate more,” according to a white paper from OpenAI.

In answer to 49% and 53% of questions on PersonQA, the company’s internal standard for gauging the precision of a model’s human knowledge, OpenAI discovered that gpt-oss-120b and gpt-oss-20b hallucinated. It is greater than OpenAI’s o4-mini model, which scored 36%, and more than three times the hallucination rate of its o1 model, which scored 16%.

Introducing the new models, according to OpenAI, the training procedures for its open models were comparable to those used for its proprietary models. The company claims that each open model operates more efficiently by using mixture-of-experts (MoE) to tap fewer parameters for any given query. According to OpenAI, the model only activates 5.1 billion parameters each token for gpt-oss-120b, which has 117 billion parameters in total.

The business also claims that high-compute reinforcement learning (RL), a post-training technique that uses massive clusters of Nvidia GPUs to teach AI models right from wrong in simulated scenarios, was used to train its open model. The o-series of models from OpenAI were also trained using this method, and they follow a similar line of reasoning, requiring more time and processing power to arrive at their conclusions.

OpenAI claims that the post-training procedure makes their open AI models excellent at driving AI agents and that they can use tools like online search or Python code execution as part of their cognitive process. OpenAI claims that their open models are text-only, which means that unlike the company’s other models, they cannot process or produce images and audio.

GPT-oss-120b and GPT-oss-20b are being made available by OpenAI under the Apache 2.0 license, which is widely regarded as one of the most permissive. With this license, businesses will be able to profit from OpenAI’s open models without having to pay the firm or get its consent.

OpenAI also claims that it will not be making the training data required to develop its open models publicly available, in contrast to completely open source products from AI labs like AI2. Given the numerous ongoing legal actions against AI model providers, such as OpenAI, alleging that these businesses improperly trained their AI models on copyrighted works, this verdict is not shocking.

In recent months, OpenAI has repeatedly postponed the distribution of its open models, in part due to safety concerns. In a white paper, OpenAI claims that in addition to its standard security procedures, it also looked into the possibility that malicious actors could modify its gpt-oss models to make them more useful for cyberattacks or the development of chemical or biological weapons.

Finally, the startup claims gpt-oss may somewhat improve biological capabilities following testing by OpenAI and outside evaluators. Nevertheless, even after fine-tuning, it found no proof that these open models could achieve the “high capability” criterion for hazard in these fields.

Even though OpenAI’s model seems to be the most advanced open model available, developers are excitedly anticipating the release of Meta’s Superintelligence Lab’s new open model and DeepSeek R2, the company’s next AI reasoning model.

With a new model that enables AI systems to interact with a realistic simulation of the actual world, Google has revealed its most recent advancement in artificial general intelligence (AGI).

As they interact with lifelike recreations of settings like warehouses, robots and self-driving cars might be trained using the Genie 3 “world model,” according to Google.

AGI, a hypothetical level of AI where a system can perform most tasks on par with humans – rather than just individual tasks like playing chess or translating languages – and possibly perform someone’s job, is argued to be achievable through world models, according to Google DeepMind, the AI division of the US technology company.

Such models would be crucial to the creation of AI agents, or systems that do tasks on their own, according to DeepMind.

“As we move closer to artificial general intelligence (AGI), and agents become more prevalent in society, we anticipate this technology to be crucial,” DeepMind stated.