Despite its name suggestion, Openai has not released an “open” model that includes access weights or numerical parameters of the brain that are often described as models – 2020’s GPT-2. Tuesday changed: The company launched two open models in two sizes, which is what Openai said, which is Open reasoning for open reasoning.
“We are excited to be able to make this model, the result of a billion-dollar research that can take AI into the hands of most people,” Openai CEO Sam Altman said in his release. “For this, we are very much hoping that this version will enable new research and create new products.” He stressed that he is “excited by the world building on the open AI stack created by the United States, based on democratic values and offering it for free for everyone, and for the broad benefit.”
Altman mocked the upcoming model two months after acknowledging China’s open model success in March, which has been “the company”In terms of historical error“When it comes to developers and builders opening up their models. But while the weights are now open, experts point out that Openai’s new model is hardly “open”. It will never give away its champion: proprietary buildings, routing mechanisms, training data, and ways to power its state-of-the-art models, including the long-awaited GPT-5, are expected to be widely released sometime this month – tightly released mid-month.
Openai targets AI builders and developers
These two new model names – gpt-oss-1220b and GPT-oss-20b – may not be possible for non-engineers, but that’s because OpenAI is turning its attention to AI builders and developers who are trying to quickly build real-world usage on their own systems. The company notes that the larger of the two models can run on a single Nvidia 80GB chips, while smaller chips are suitable for consumer hardware on MAC laptops, such as Mac laptops.
Openai co-founder and president Greg Brockman admitted in a news trailer that it has been “a long time” since the company released its open model. He added that this is “we think complementary to other products we release”, and Openai’s proprietary model: “Together, it can really accelerate our mission to ensure that the API benefits all humans.”
Openai said the new models performed well in terms of reasoning benchmarks that have become key measurements of AI performance, and models from OpenAI, Anthropic, Hotlink, Google, (/Hotlink), (/Hotlink), DeepSeek are fiercely competing with their capabilities to solve their capabilities to solve multi-step logic, code, code generate complex problems and complex problem solutions. Since the open-source DeepSeek R1 shocked the industry with its reasoning capabilities in January with much lower costs, many other Chinese models have followed suit, including Alibaba’s Qwen and Moonshot AI’s Kimi models. Although Openai said at a press conference that the new open weight model is an active effort for what users want, it is obviously also a strategic response to improving open source competition.
It is worth noting that Openai refuses to benchmark its new open weight models for Chinese open source systems such as DeepSeek or Qwen, although these models have performed well on key inference benchmarks recently. The company said in a press conference that it has confidence in its model and leaves it to others in the AI community to further test and “make up our minds.”
Avoid intellectual property leaks
OpenAI’s new open weight model is built using a mixture of Experts (MOE) architectures, where the system activates only “experts” or subnetworks, which requires specific inputs, rather than using the entire model for each query. Dylan Patel, founder of semi-analytics in a post about x Prior to release, Openai trained the model using only public components of the architecture, meaning the open source community is already familiar with the artifacts used. He stressed that this was a deliberate choice – avoiding any proprietary training technology or architectural innovation, Openai could issue truly useful models without actually leaking any intellectual property that could power GPT-4O, such as GPT-4O.
For example, in the model card that accompanied the release, OpenAI confirms that the model uses a mixture of Experts (MOE) structures, with 12 active experts, but it does not describe the routing mechanism for the critical and proprietary parts of the architecture.
“You want to minimize business risk, but you (also) want to be most useful to the public,” former Google Aleksa Gordic Deep state Researchers, tell wealth, Added that the company likes it Yuan Likewise, no proprietary information is included.
“They minimize IP leaks and eliminate any risks to their core business while sharing useful artifacts, which will enable the startup ecosystem and developers to do it,” he said. “That’s the best they can do by definition, given these two relative goals.”
