OpenAI Launches GPT-OSS: Free AI Models that Can Run On Your Laptop

OpenAI has officially taken a major step toward openness with the release of its new open-source weight model, GPT-OSS This means you can download and modify it for free, and surprisingly, it can even run on your laptop. The democratization of AI is unfolding in a whole new way.

OpenAI has launched GPT-OSS, a model that comes in two different versions, each adapted for different hardware and usage scenarios. The larger version has 120 billion parameters and powerful performance, and can run directly on an Nvidia GPU, with performance comparable to OpenAI's existing flagship model, the o4-mini. On the other hand, the smaller 20 billion parameter version is uniquely optimized with an absolute memory requirement of 16GB, and can even work on your regular laptop with performance similar to the o3-mini! Both versions have been released on mainstream platforms such as Hugging Face, Databricks, Azure, and AWS, and use the permissive Apache 2.0 license, allowing users to freely modify and commercialize them according to their needs. Whether you are a large enterprise or an individual developer, you can easily use this model to create your own AI solution, truly achieving a combination of high performance and flexibility.

 

Whether you're building enterprise applications or conducting experiments on personal devices, you can find a version that's right for you. More importantly, both models have been released under the Apache 2.0 open source license on major platforms like Hugging Face, Databricks, Azure, and AWS. This allows you to freely modify them, use them commercially, and even seamlessly integrate them into your own products, truly enabling flexible and controllable application development.

This isn't just the release of a new model; it marks a significant shift in OpenAI's open-source strategy. Since GPT-2, OpenAI hasn't publicly released any open-weight models—GPT-3, GPT-4, and the O series have all remained closed. In other words, GPT-OSS is the first time the core weights of the GPT series have been directly accessible to the public. Until earlier this year, OpenAI CEO Sam Altman insisted that they wouldn't release open models for security reasons. However, with the release of open-source models like LLaMA, Mistral, Gemma, and DeepSeek, the needs of the developer community have rapidly evolved, with a growing desire to run models locally and flexibly deploy them, prompting a strategic shift in OpenAI's approach.

As DeepSeek garnered widespread attention, Altman's attitude shifted significantly. He frankly admitted that his previous refusal to release his own open models was "on the wrong side of history." This wasn't just a moment of self-reflection; it signaled OpenAI's realization that continuing with a closed strategy would be difficult to withstand the rapidly rising tide of open source. The emergence of GPT-OSS emerged precisely in this context. It not only addressed previous hesitations but also clearly signaled OpenAI's new stance on openness amidst fierce competition.

OpenAI went beyond "engaging in openness" this time, it was ultimately looking to reassert its technological leadership in the AI community through GPT-OSS. GPT-OSS is different from traditional open source models, it has what can be considered, two usages: can be deployed independently in local environments for private uses, and can be bundled in existing services via the OpenAI API, enabling a variety of tasks that include: inference, web browsers, writing code, and agent controls.It serves as both a powerful tool for developers and a critical bridge connecting the OpenAI technology ecosystem.

OpenAI researcher Chris Cook admitted at the press conference, "Many people may not realize that most of our customers are already using various open models." This statement reveals a crucial fact: open source technology is gaining increasing market acceptance, and OpenAI is actively embracing this proven trend.

He further stated, "We want to fill this gap and enable users to fully leverage our technology." This isn't just a product launch, but rather a reshaping of the ecosystem—OpenAI is committed to attracting users who already use open models back to its platform and ecosystem.

However, openness also presents challenges and risks. Due to long-standing security concerns, OpenAI is proceeding with extreme caution with the release of GPT-OSS. Officials emphasized that this is their most rigorously tested model to date, having undergone multiple rounds of internal review and involving external security agencies in assessments to ensure it won't be misused for sensitive areas such as cybersecurity threats or biological weapons.

To make the model's behavior more transparent and controllable, OpenAI introduced a "thought chain" mechanism in GPT-OSS. Thus, when the model claims it has an answer, it is making its rationale and reasoning visible. This not only makes the results easier to understand, but also provides a strong safeguard against potentially misleading or anomalous outputs.

Importantly, GPT-OSS outputs only plain text so there is no possibility for covert dissemination via images or code. Furthermore, as with previous OpenAI models, its training data is not publicly available, in part to protect the model from reverse engineering and data poisoning.

This demonstrates that despite GPT-OSS's open nature, OpenAI is still striving to strike a balance between security and freedom.

OpenAI has not yet released detailed benchmarks comparing GPT-OSS with other mainstream open models, such as Llama, DeepSeek, or Google's Gemma, leaving much room for improvement.

Despite this, OpenAI stated that both versions of GPT-OSS performed exceptionally well in tests such as programming tasks and the highly anticipated "The Last Human Exam," demonstrating that their performance has reached industry-leading levels. Co-founder Greg Brockman also praised this achievement, stating, "These models are exceptional, and the team's efforts are admirable."

This has undoubtedly garnered increased attention and anticipation for GPT-OSS within the open model community, creating new dynamics in the future competitive landscape.

OpenAI has not yet announced a specific release schedule for subsequent versions of GPT-OSS, but they hope that the model will be widely adopted by small developers and businesses seeking greater control over how their data is used.

As co-founder Greg Brockman said, "We firmly believe that innovation will emerge naturally when the technical barriers to entry are lowered. When people have the freedom to explore and adapt, they will create many surprising and exciting results."

This not only reflects the core value of the open model, but also demonstrates OpenAI's firm confidence in the continued prosperity of the future technology ecosystem—openness brings infinite possibilities, and innovation will continue to emerge.