OpenAI has recently made a significant shift in its approach to artificial intelligence by releasing two new open-weight models, gpt-oss-120B and gpt-oss-20B. This marks the first time in six years that the company has offered such models, which can now run directly on personal devices such as laptops and be fine-tuned for a variety of applications. This release is particularly noteworthy as it comes after multiple delays attributed to safety concerns.

In a blog post, OpenAI expressed their excitement about providing these best-in-class open models. They aim to empower everyone from individual developers to large organizations and government entities to run and customize AI on their own infrastructure. The timing of this launch is particularly interesting, following the earlier release of DeepSeek’s cost-effective, open-weight R1 model, which may have influenced OpenAI’s strategy to diversify away from closed proprietary models that have dominated their offerings since the 2019 launch of GPT-2.

Alongside these models comes the anticipation of the GPT-5 model that OpenAI is expected to release shortly. So what do we know about the new gpt-oss models? The gpt-oss-120B model, boasting an impressive 117 billion parameters, is capable of running on a single 80GB GPU, while its smaller counterpart, the gpt-oss-20B, can be deployed even on a laptop with merely 16GB of memory.

Both models have been released under the Apache 2.0 license, allowing developers to download and host them freely on platforms like Hugging Face. Microsoft is also adapting a GPU-optimized version of the gpt-oss-20B model for Windows devices, further broadening the reach and accessibility of these models.

The sheer number of parameters in an AI model can often correlate with its problem-solving capabilities. By conventional understanding, models with a higher parameter count generally exhibit better performance. OpenAI, however, claims to have made these new models more efficient using a mixture-of-experts (MoE) technique, which DeepSeek also employs. This method enhances energy efficiency and reduces computational costs by activating only a small fraction of the model’s parameters for specific tasks.

In addition to this, OpenAI has employed grouped multi-query attention to optimize inference and memory efficiency, which diverges from the multi-head latent attention technique seen in DeepSeek’s V2 model. This attention mechanism is particularly important for enhancing performance in extensive applications where quick and efficient response times matter.

Interestingly, the gpt-oss models support a maximum context window of 128,000 tokens, a notable feature that expands the potential for context-rich interactions, further enhancing their utility in various applications.

As for performance comparisons, the gpt-oss-120B model has been reported to match the performance levels of o4-mini, OpenAI’s most cutting-edge model to date. This indicates that even though the open-weight models are positioned as more accessible options, they do not compromise on performance, making them viable alternatives for businesses and individual developers alike.

The release of these open-weight models signifies a critical moment in AI history as it opens the door for broader participation in the AI landscape. By allowing developers to customize models according to their specific needs and run them on local infrastructures, OpenAI encourages innovation and reduces dependency on cloud-based solutions. This move has vast implications for businesses looking to leverage AI tools tailored precisely to their operational challenges.

However, questions remain regarding the true openness of these models, stirring discussions in the AI community about the balance between accessibility and control over powerful AI systems. As OpenAI champions this new direction, stakeholders will be watching closely, hoping it catalyzes a wave of advancements while also emphasizing the importance of responsible AI development.