OpenAI has unexpectedly shaken up the artificial intelligence landscape with the launch of GPT-OSS , its first major open-source model in over five years. This development marks a departure from the company's recent strategy, which had focused primarily on proprietary models and closed services. With GPT-OSS, OpenAI returns to its roots, opening the door for anyone or any organization to access, run, and customize a powerful language model without relying on its cloud platform or internet connections.
The new model is available in two versions: one with 120.000 billion parameters and a lighter one with 20.000 billion . The more powerful variant can run on a single professional GPU with at least 80 GB of memory, while the smaller version is geared towards conventional desktop and laptop computers with 16 GB of RAM, significantly democratizing access to advanced AI capabilities . Both versions can be downloaded for free from platforms like Hugging Face and deployed on popular services such as Azure or AWS.
Technical characteristics and technology used
With GPT-OSS , OpenAI has introduced two models designed with a Transformer-based architecture and mixture-of-experts (MoE) . This optimizes memory usage and latency by activating only a select fraction of experts per input token, increasing efficiency without sacrificing power. The 120B edition uses 36 blocks and activates 5.100 billion parameters per token, while the 20B edition uses 24 blocks and activates 3.600 billion, making it suitable for more affordable hardware.
Both models were trained primarily on English-language data from STEM fields, programming, and general knowledge , including an additional phase of supervised tuning and reinforcement to improve their alignment with human instructions . Their capabilities include chain-of-thought reasoning , the ability to break down intermediate steps before providing a response, and the use of external tools such as web browsing and Python code execution.
Availability, licensing and integration options
One of the major advantages of GPT-OSS is its Apache 2.0 license , which eliminates many of the usual restrictions of other proprietary models: it allows for commercial use, redistribution, and integration into all types of projects, from small startups to large public organizations or R&D initiatives. The model weights are available in MXFP4 format and include reference implementations in PyTorch and Apple Metal, as well as full compatibility with tools such as Ollama, llama.cpp, LM Studio, and vLLM , facilitating its use in both professional and home environments.
Installing and deploying GPT-OSS-20B is particularly easy thanks to utilities like LM Studio, which allows users with no programming experience to download and test the model locally using a graphical interface. Furthermore, OpenAI has published recommendations and detailed documentation for fine-tuning and adapting the models, as well as ready-made integrations for cloud and on-premises platforms.
Safety, testing and responsible approach
The launch of GPT-OSS has focused not only on openness but also on security and the responsible use of the models. OpenAI has incorporated new review and validation mechanisms to prevent risks associated with misuse, especially in sensitive areas such as cybersecurity and the creation of malicious agents. To this end, it has implemented an adversarial testing methodology (red teaming), with external reviews by independent experts and protocols to simulate extreme exploitation scenarios.
Alongside the release of the models, OpenAI has launched a network teaming challenge with financial rewards for those who identify vulnerabilities or problematic behavior, seeking active collaboration from the global community in risk detection. Furthermore, GPT-OSS exposes its reasoning chain , facilitating the auditing of responses and monitoring to prevent bias or critical errors.
Performance and comparison with other open models
In academic benchmarks and reasoning tests , GPT-OSS-120B performs on par with, and even surpasses in some areas, recent proprietary OpenAI models like o4-mini, and outperforms others such as DeepSeek R1 or the Llama series in certain contexts. The 20B version, optimized for home devices, maintains an excellent balance between efficiency and performance , with highly competitive scores in math, coding, and health tasks.
OpenAI warns that these open models may be more prone to "hallucinations" (factual errors) compared to their cloud-based counterparts, given the differences in training and monitoring. However, users now have the option of using high-level models without needing online services.
