Openai launches “open” thinking models

Openai announced on Tuesday the launch of logical thinking models of artificial intelligence open weight with similar capabilities For the O.. Both are free to download from the online developer platform EmbroideryThe company said, the models describe as “the latest technology” when measured by several criteria to compare the open models.

Models come in two-sizes: the larger and more capable GPT -SS-120B model that can be operated on a single NVIDIA graphics processing unit, and the GPT -SS-20B model is lighter weighing can work on a laptop of consumer with 16 GB memory.

The launch is the first “open” language model for Openai since then GPT-2Which was issued more than five years ago.

In briefing, Openai said that its open models will be able to send complex inquiries to Artificial intelligence models in the cloud, As mentioned previously Techcrunch. This means that if Openai’s open model is not capable of a specific task, such as processing an image, developers can connect the open model to one of the closed models that are the most capable of the company.

Although Openai Open Sourted AI models are in its early days, the company generally preferred the closed source development approach. The recent Openai strategy helped build great works that sell access to artificial intelligence models via API for institutions and developers.

However, CEO Sam Al -Tamman said in January he believed that Openai was “on the wrong side of history” when it comes to opening the sources of his techniques. Today, the company faces increasing pressure from Chinese AI Labs – including Deepseek, Alibaba’s Qwen and Moonshot AI – which has developed many of the world’s most capable open models. (While Meta previously dominated the open AI space, Llama AI models for the company have Backwardness Last year.)

In July, the Trump administration also urged us to developers Open source more technology To enhance global adoption of artificial intelligence with American values.

With the GPT -SS version, Openai hopes that the curry would prefer with the developers and the Trump administration alike, both of whom witnessed the Chinese AI laboratories rising in the open source space.

“By returning to when we started in 2015, Openai’s mission is to ensure that AGI benefits all humanity,” Altman said in a joint statement with Techcrunch. “To achieve this purpose, we are enthusiastic about the world to build on the Mixture of Amnesty International Open in the United States, based on democratic values, and are available for free for all and to the tendencies.”

How models were implemented

Openai aims to make it Open Open Leader Among other models of open -weight artificial intelligence, the company claims it did.

On Codeforcees (with tools), competitive coding test, GPT -SS-120B and GPT-SS-20B 2622 and 2516, respectively, surpasses R1 Deepseek during the performance of O3 and O4-MINI.

In the last Humanity Exam (Hle), a difficult test of collective questions through a variety of topics (with tools), GPT -SS-120B and GPT-SS-20B by 19 % and 17.3 %, respectively. Likewise, this weak performance O3 but surpasses open models of Deepseek and QWEN.

It is worth noting that OpenAi open models calm down much more than the latest models of male thinking, O3 and O4-MINI.

Halousa was Get more severe In the latest models of artificial intelligence from Openai, the company said previously it does not fully understand the reason. In a white sheet, Openai says that this is “expected, because smaller models have a world less than large border models and tend to hallucinations more.”

Openai found that GPT -SS-120B and GPT-SS-20B in response to 49 % and 53 % of questions on PersonQa, the company’s inner criterion for measuring the accuracy of the form of the model about people. This is more than three times the hallucinations rate for the OPNAI O1 model, which recorded 16 %, and the highest O4-MINI model, which recorded 36 %.

Training new models

Openai says its open models have been trained with operations similar to their royal models. The company says that every open model benefits from experience (MEE) to take advantage of a fewer parameters for any specific question, which makes it work more efficiently. For GPT -SS-120B, which includes 117 billion parameters, Openai says the model is only 5.1 billion teachers for each symbol.

The company also says its open style has been trained using high computing Reinforce learning (RL) The post -training process to teach the correct artificial intelligence models from an error in simulation environments using large groups of NVIDIA graphics processing units. This was also used to train the Openai series of models, and the open models have a similar detailed process as it takes additional time and calculation resources to work through its answers.

As a result of the post -training process, Openai says that the open AI models outperform the operation of artificial intelligence agents and are able to call tools such as searching on the web or implement the Python icon as part of the idea series process. However, Openai says that its open models are only a text, which means that they will not be able to process or create images and sound like other company models.

OpenAi release GPT -SS-120B and GPT-SSS-20B under APache 2.0 license, which is generally considered one of the most names. This license will allow institutions to transfer open models for Openai without the need to pay or obtain permission from the company.

However, unlike the fully open sources of AI Labs, such as AI2, Openai says it will not issue training data used to create their open models. This decision is not surprising, given that many active lawsuits against the artificial intelligence model providers, including Openai, claimed that these companies have trained inappropriately trained artificial intelligence models on copyright -protected business.

Openai delay Issuing its open models numerous Times in recent months, partially to address safety concerns. Besides the model safety policies of the company, Openai says in a white paper it also achieves whether bad actors can adjust the GPT -SS models to be more useful in electronic attacks or create biological or chemical weapons.

After the test from Openai and Thred-theths-thethships, the company says that GPT-SS may increase in a marginal way of biological capabilities. However, he did not find evidence that these open models can reach the threshold of their “high ability” at risk in these areas, even after control.

Although the Openai model appears to be among the latest among open models, developers are eagerly awaiting Deepseek R2, a model of thinking from artificial intelligence, as well as a new open model of the Meta’s SuperIntellent Lab.