Insights on OpenAI's Latest Open Weight AI Models: Exploring Cost, Speed, and Availability for Public Use
In a significant move, OpenAI has released two new open source and open weight AI models, named gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 license. These models, which offer substantial innovation advantages, aim to address some of the risks associated with open source AI models.
The key risks of open source AI models, as highlighted by various experts, include security vulnerabilities, misuse potential, legal complications, and challenges in governance and oversight. Open source AI models and dependencies can contain unpatched vulnerabilities that attackers might exploit, leading to data poisoning attacks, introducing backdoors, or stealing sensitive data.
To mitigate these risks, OpenAI has reinforced mechanisms to reject attempted prompt injection in the gpt-oss models. They have also taken steps to ensure the models cannot be used for "high capability" malicious purposes. Notably, OpenAI has removed harmful data related to Chemical, Biological, Radiological, and Nuclear (CBRN) from the models.
OpenAI has also acknowledged the risk of open source and open weight AI models empowering attackers. To counter this, they have implemented a mixture of experts (MoE) approach for faster inference and less expensive pre-training. This approach, combined with rigorous lifecycle security practices, is designed to create resilient AI systems while retaining many open source benefits.
The gpt-oss models, available on platforms like Hugging Face, Azure, AWS, and Databricks, offer price performance that is 10 times that of AWS's Gemma counterpart and 18 times that of DeepSeek-R1, according to AWS claims. The smaller model, gpt-oss-20b, has 21 billion parameters and can be run with 16GB of memory, making it accessible for personal computers and even phones.
The larger model, gpt-oss-120b, contains 117 billion parameters and can be run on a single Nvidia A100 GPU with 80GB of memory. In benchmarks, gpt-oss outperformed o4-mini across health and expert questioning benchmarks but fell behind slightly on code completion and math tasks.
OpenAI has also launched a new AI agent for software engineering called Codex. This development further expands OpenAI's portfolio, demonstrating their commitment to advancing AI technology while addressing the associated risks.
To encourage further scrutiny and ensure the models' safety, OpenAI has opened a red teaming challenge for gpt-oss-20b, with a maximum prize of $500,000 for evidence of exploits. This proactive approach underscores OpenAI's commitment to maintaining the highest standards of security and safety in their AI models.
The field of open source and open weight AI models is evolving towards hybrid approaches that balance openness with control and visibility. OpenAI's release of the gpt-oss models and their commitment to addressing associated risks are significant steps in this direction.
- In an attempt to safeguard the gpt-oss models from potential cybersecurity threats, OpenAI has reinforced mechanisms to resist prompt injection and implemented a mixture of experts (MoE) approach for improved inference and pre-training, aiming to create resilient AI systems.
- Recognizing the risks posed by open source and open weight AI models, OpenAI has taken proactive steps, such as removing harmful data related to Chemical, Biological, Radiological, and Nuclear (CBRN) from the models, and has even launched a red teaming challenge for gpt-oss-20b, offering a $500,000 prize for evidence of exploits, to encourage further scrutiny and ensure the models' safety.
