OpenAI's Latest Generative AI Models Are Here!

In a groundbreaking announcement, OpenAI has just revealed two new generative AI models—gpt-oss-120b and gpt-oss-20b. While they aren't the much-anticipated GPT-5, these open-weight models are the first since GPT-2, and they're turning heads for all the right reasons! These models are designed to be downloaded and operated on your own hardware, offering features such as simulated reasoning, tool use, and deep customization.

A Glimpse at the New Models

The release features two variants, catering to different hardware capabilities. The gpt-oss-120b boasts a whopping 117 billion parameters, while the gpt-oss-20b has 21 billion parameters. Both models include a configurable chain of thought (CoT) setting that allows users to optimize their processing power: - **Low settings** for speed and minimal resource use. - **High settings** for top-tier output quality.

Performance and Accessibility

The gpt-oss-20b can run on a standard consumer machine with 16GB of memory, making it accessible for many users. Meanwhile, the larger gpt-oss-120b requires about 80GB of memory, ideal for high-performance setups such as the Nvidia H100 GPU. Notably, both models feature an expansive context window of 128,000 tokens, providing ample space for complex tasks.

Robust Performance Compared to Proprietary Models

Users can expect performance comparable to OpenAI's top cloud-based models. Initial benchmarks show the gpt-oss-120b performing closely to proprietary models, especially in math and coding tasks. However, it's worth mentioning that it trails Google’s advanced Gemini Deep Think model.

Open-Source with Limitations

OpenAI emphasizes that these new models are not intended to replace its proprietary versions. gpt-oss-120b and gpt-oss-20b are limited to text-only capabilities, lacking multimodality features. However, they provide a lower-latency, customizable alternative for those wary of relying on cloud-stored data.

A Focus on Safety and Control

To maintain ethical standards, OpenAI has integrated controls to limit malicious use. Interestingly, during testing, the models were challenged to behave poorly but maintained a low-quality output for malicious activities, thanks to the company’s alignment framework.

Get Started Today!

Eager to explore these models? You can download gpt-oss-120b and gpt-oss-20b today from Hugging Face. Plus, check out GitHub for additional resources. For the tech-savvy, OpenAI has also made model cards and detailed research materials available, so dive in and innovate!