The initial reactions to the Open Source GPT-AS models are very varied and mixed
Do you want smarter information in your reception box? Sign up for our weekly newsletters to obtain only what matters for business managers, data and security managers. Subscribe now
The long -awaited return of Openai to “the opening” of its namesake occurred yesterday with the release of two new models of great language (LLMS): GPT-OS-150B and GPT-OS-20B.
But despite the realization of technical benchmarks tied with the other powerful offers of OPENAI owners, the AI developer and the wider user community So far, the answer has been everywhere on the map. If this version was a film in the first and in the course of the tomato note rot, we would consider a split of almost 50%, on the basis of my observations.
First of all, Openai published these two new text language models only (no generation of images or analysis) Both under the Apache 2.0 Open Source Permissive license – The first time since 2019 (before the Chatppt) that the company did it with a model of advanced language.
THE The whole chatgpt era in the past 2.7 years has been fueled by proprietary or closed source models until nowThose who Openai checked and users had to pay to access (or use a free level subject to limits), with limited personalization and no way to execute them offline or on private computer equipment.
The AI scale reached its limits
Electricity ceilings, increase in token costs and inference delays restart the AI company. Join our exclusive fair to discover how best the teams are:
- Transform energy into a strategic advantage
- Effective inference architecting for real debit gains
- Unlock a competitive return on investment with sustainable AI systems
Secure your place to stay in advance::
But everything has changed thanks to the release of the pair of GPT-OS models yesterday, a larger and more powerful for use on a single NVIDIA H100 GPU at a data center or a farm of small or medium or medium servers, and an even smaller PC on a single laptop of consumption or office like the type in your home office.
Of course, the models being so new, it took several hours for the AI power user community to execute it independently and to test them on their own benchmarks (measurements) and tasks.
And Now we get a wave of comments ranging from optimistic enthusiasm on the potential of these new powerful, free and effective models to an underlying current of dissatisfaction and dismay of what certain users consider as important problems and limitationsespecially compared to the license wave similar to Apache 2.0 LLMS Open Source Powerful of Chinese startups (which can also be taken, personalized, execute locally on American equipment free of charge by American companies or businesses everywhere else in the world).
High benchmarks, but still behind Chinese open source chiefs
Intelligence references place the GPT-AS models before most of the American open-source offers. According to an artificial artificial analysis of the Artificial Information Analysis Society of the independent AI, GPT-AS-15B is “the most intelligent American weight model”, although it does not become Chinese heavyweights like Deepseek R1 and Qwen3 235b.
“On reflection, that’s all they have done. Mogged on references,” wrote Deepseek “Stan” @tortaxestex self -proclaimed. “No good derivative model will be formed … no new usecases created … A sterile claim to boast rights.”
This skepticism is taken up by the pseudonym researcher of the Open Source Teknium (@ teknium1), co-founder of the AI Open Source Rival model provider, a research on the Chinese version, which called the version “a legitimate hamburger”, on X, and predicts that a Chinese model will soon eclipse it. “Overall, very disappointed and I legitimately opened up to mind,” they wrote.
Banc-maxxing on mathematics and code to the detriment of writing?
Other criticisms focused on the Apparent narrow utility of GPT-AS models.
The influencer of the AI “Lisan al Gaib (@ scaling01)” noted that the models excellent in mathematics and in coding but “completely lack taste and common sense”. He added: “So it’s just a mathematical model?”
In creative writing tests, some users have found the model injecting equations in poetic outings. “This is what happens when you benchmarkmax,” said Teknium, sharing a screenshot where the model added a full formula in mid-pole.
And @Kalomaze, researcher of the decentralized IA model training company, Prime Intellect, wrote that “GPT-OS-120B knows less about the world than what a good 32B does. I probably wanted to avoid copyright problems, so they probably have a preterity on the majority synth. Quite devastating stuff ”
The former Googler and developer of independent AI Kyle Corbitt agreed that the pair of GPT-AS models seemed to have been trained mainly on synthetic data-that is to say data generated by a model of AI specifically specifically for training purposes-which makes it “extremely thorny”.
It is “great in the tasks on which he is formed, really bad in everything else,” wrote Corbitt, that is to say, Super on coding and mathematical problems, and bad to more linguistic tasks such as creative writing or generation of reports.
In other words, the accusation is that Optai deliberately trained the model on more synthetic data than the facts and figures of the real world to avoid using data protected by copyright on websites and other benchmarks that he does not have or has no license to use, which is something of which many other Gen AI societies have been accused in the past and are confronted with current prosecution.
Others have assumed that OPENAI may have formed the model on mainly synthetic data to avoid security and safety problems, which has led to worse quality than if it had been trained on more real (and probably protected) data.
Regarding third -party reference results
In addition, the evaluation of models on third -party comparative analysis tests has proven to be concerned about the measures in the eyes of certain users.
SPE: near the bottom of open peers models, This indicates resistance to follow user requests and a lack of care, potentially to the detriment of the provision of precise information.
In the polyglot assessment to help, GPT-OS-120B marked only 41.8%in multilingual-FAR reasoning below competitors like Kimi-K2 (59.1%) and Deepseek-R1 (56.9%).
Some users have also said that their tests indicated that the model was strangely resistant to the generation of criticism from China or Russia, a contrast to its treatment from the United States and the EU, raising questions about biases and filtering training data.
Other experts have applauded the version and what it signals for us open source
To be fair, the whole comment is not negative. The software engineer and the supervisor of the ENA ENS Simon Willison called the “really impressive” version on X, developing in a blog article on The efficiency and ability of models to reach parity with O3-Mini and O4-Mini of Openai owner models.
He praised their strong performance on reasoning and heavy references of stems, and praised the new “Harmony” prompt model format – which offers developers more structured terms to guide model responses – and support for the use of third -party tools as significant contributions.
In a long post X, Clem Delangue, CEO and co-founder of the AI code sharing and the face of the open source community, encouraged users not to rush to judgment, stressing that the inference for these models is complex, and the first problems could be due to the instability of infrastructure and an insufficient optimization between accommodation suppliers.
“The power of the opening is that there is no cheating,” wrote Delangue. “We will discover all the forces and limits … gradually.”
Even more prudent was the Wharton School of Business by Professor Ethan Mollick of the University of Pennsylvania, which wrote on X that “the United States probably have the main models of open weight (or near it)”, but wondered if it was a single Off of Openai. “The advance will evaporate quickly while others are caught up”, “ He noted, adding that it is not clear what OpenAi incentives have to keep the models up to date.
Nathan Lambert, a main researcher of AI at the Open Source Rival Lab Allen Institute for IA (AI2) and the commentator, praised the symbolic meaning of the version on his interconnects blog, calling him “A phenomenal step for the open ecosystem, especially for the West and its allies, That the best known brand in AI space has returned to the release of models openly. »»
But he warned against X that GPT-Ass “It is unlikely that the slowdown (the AI team of the Chinese electronic commerce giant Aliba) has slowed down) qwen”, ” Quoting his conviviality, his performances and his variety.
He argued that the release marks a significant change in the United States to open models, but this OpenAi still has a “long return path” to catch up in practice.
A divided verdict
The verdict, for the moment, is divided.
OPENAI GPT-AS models are a benchmark in terms of license and accessibility.
But while benchmarks seem solid, the “vibrations” of the real world – as many users describe it – prove to be less convincing.
That developers can build solid applications and derivatives above GPT-OS will determine whether release is recalled as a breakthrough or a Blip.
https://venturebeat.com/wp-content/uploads/2025/08/ChatGPT-Image-Aug-6-2025-02_39_56-PM.png?w=1024?w=1200&strip=all
About The Author
