OpenAI Introduces GPT-OSS: An Open-Weight Reasoning Models Optimized for Laptops

OpenAI has unveiled a new open-weight model named GPT-OSS, available for free download and customisation. This model comes in two versions: one with 120 billion parameters and another with 20 billion. The larger version can operate on a single Nvidia GPU, matching the performance of OpenAI's o4-mini model. Meanwhile, the smaller variant is comparable to the o3-mini and requires only 16GB of memory.

The release of GPT-OSS marks OpenAI's first open-weight model in over six years, predating ChatGPT. CEO Sam Altman previously cited safety concerns as a reason for not releasing such models earlier. However, developers have increasingly turned to open models due to their affordability and flexibility. In January, Altman acknowledged that OpenAI had been "on the wrong side of history" by not offering its own open models.

OpenAI's New Model Features

Both versions of GPT-OSS are now accessible through platforms like Hugging Face, Databricks, Azure, and AWS under the Apache 2.0 license. This licensing allows for extensive modification for commercial use. The model is designed to perform reasoning tasks, browse the web, write code, and operate agents via OpenAI's existing APIs.

Chris Cook, an OpenAI researcher, highlighted during a media briefing that many customers already utilise open models extensively. "We wanted to plug that gap and allow them to use our technology across the board," he stated.

Focus on Safety and Testing

OpenAI emphasises that GPT-OSS has undergone rigorous testing with external safety firms to ensure it does not pose risks in areas such as cybersecurity or biological weapons. The model's chain of thought is visible to monitor potential misbehaviour or misuse. Despite its capabilities, GPT-OSS's training data remains undisclosed.

The company has not provided benchmarks comparing GPT-OSS with other open models like Llama or Google's Gemma. However, both variants perform similarly to OpenAI's closed reasoning models on coding tasks and tests like Humanity's Last Exam.

Encouraging Innovation

OpenAI cofounder Greg Brockman expressed optimism about the potential uses of GPT-OSS by smaller developers and companies seeking more control over their data usage. "We've always believed that if you lower the barrier to access, then innovation just goes up," Brockman remarked.

While OpenAI hasn't committed to a specific release schedule for future versions of GPT-OSS, it hopes this model will inspire unexpected innovations from users who explore its capabilities.